Recognition of Handwritten Arabic Characters by ...

International Review on

Computers and Software (IRECOS)

PART

A

Contents FDTD Simulation Method Using Double Signals for Electromagnetic Field: Determination of the Tissue Biological Characteristics by Mondher Chaoui, Moez Ketata, Mongi Lahiani, Hamadi Ghariani

874

Recommendation Systems Through Semantic Web Usage Mining by Mehrdad Jalali, Norwati Mustapha

882

Prototyping Mobile Graphical Authentication by Martin Mihajlov, Borka Jerman Blažič, Dragan Tevdovski

887

Some Open Problems in Multimedia Digital Fingerprinting by Song Tang, Li Liu

894

QoS Correction in IMS Networks by M. Errais, B. Raouyane, M. Bellafkih, M. Ramdani

898

A Strawberry Disease Image Retrieval Method Inosculating Color and Textural Features by Jian Song

908

A Robust Audio Watermarking Scheme Based on SVD and MDCT by Ping Su, Haidong Shi

913

Recognition of Handwritten Arabic Characters by Using Reduction Techniques by Salah M. Al-Saleh, Salameh A. Mjlae, Salim A. Alkhawaldeh

918

SVM-kNN Fusion in Vocabulary Tree Method for Specific Object Recognition by Amir Azizi, Sattar Mirzakuchaki

924

Energy Efficient Routing Protocols for Wireless Sensor Networks: a Survey by K. S. Shivaprakasha, Muralidhar Kulkarni

929

Efficiency of MDB Routing Algorithm Over DB Routing Algorithm in Point-Point Networks by S. Anuradha, G. Raghu Ram, T. Bhaskara Reddy, J. Chakradhar

944

Real Time “2CRT” Architecture and Platform for the Experimentation of Telecommunication Terminals According to the Manhattan Mobility by M. El Bakkali, H. Medromi

950

A Bandwidth Request Competition Mechanism in WiMAX System by Jianbo Ye, Qingbo Zhang

956

An Inter-Cluster Cooperative Nodes Selection Scheme Based on Blind Channel Estimation by Xiaoqiang Zhong, Baiqing Zhou

960

A Suitable Architecture and Deployment Considerations for Shallow Water Acoustic Sensor Networks by V. Hadidi, R. Javidan, M. Keshtgary, A. Hadidi

965

Creating Barcode Using Fractal Theory by K. Kiani, E. Kosari, M. R. Simard

979

(continued)

Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Model Proving of Urban Traffic Control Using Neuro Petri Nets and Fuzzy Logic by Rishi Asthana, Nilu Jyothi Ahuja, Manuj Darbari

983

A New Evaluation Strategy Based on SVM for Supply Chain Finance by Wenfang Sun, Jindong Zhang, Xilong Qu

988

Leveraging Historical Assessment Records for Constructing Concept Maps by M. Al-Sarem, M. Bellafkih, M. Ramdani

994

Object Oriented Database Applying Study on ISO 9001:2000 System by K. Khoualdi, T. Alghamdi

1001

A Hybrid Particle Swarm Algorithm for Job Shop Scheduling Problem by Gongfa Li, Yuesheng Gu, Hegen Xiong, Jianyi Kong, Siqiang Xu

1006

Customer Classification Based on Data Mining by Peishuai Chen, Chonghuan Xu, Fuguang Bao

1013

Fuzzy Support Vector Machines Based on Dual Split Off Sample Space by Hongyan Pan

1018

Product Line Production Planning Model Based on Genetic Algorithm by Guozhang Jiang, Yuesheng Gu, Jianyi Kong, Gongfa Li, Liangxi Xie

1023

Moderating Effect Analysis of Relationship Length in Relationship Benefits Using Data Mining by Qingmin Kong, Xiuqing Liang

1028

Human Action Recognition Using Local Features in Bag of Keypoints Paradigm by Reyhaneh Rashidi, Javad Haddadnia

1036

Virtual Hand Modeling and Simulation Based on Unity 3D by Lam Meng Chun, Haslina Arshad

1044

Forecasting Job Cycle Time in a Wafer Fabrication Factory by the FPCA-FBPN Approach by Toly Chen

1050

A LED Landscape Lamp Automatic Real-Time Control Schema Based on Fuzzy Neural Network by Daiyun Weng, Li Yang

1055

Hardware Designs and Architectures for Projective Montgomery ECC Over GF (p) Benefiting from Mapping Elliptic Curve Computations to Different Degrees of Parallelism by Mohammad Al-Khatib, Azmi Jaafar, Zuriati Zukarnain, Mohammad Rushdan

1059

A New Digital Technique in Predicting Implant Size for Pre-Operative Planning in Total Hip Arthroplasty by Azrulhizam Shapi’i, Riza Sulaiman

1071

Improvement of Wood Ultrasonic CT Images by Using Time of Flight Data Normalization by Honghui Fan, Hongjin Zhu, Guangping Zhu, Xiaojie Liu

1079

(continued on Part B)


International Review on Computers and Software (I.RE.CO.S.), Vol. 6, N. 6 November 2011

FDTD Simulation Method Using Double Signals for Electromagnetic Field: Determination of the Tissue Biological Characteristics Mondher Chaoui, Moez Ketata, Mongi Lahiani, Hamadi Ghariani

Abstract – This paper describes the method to determine the frequency of an UWB (Ultra Wide Band) wave at which the wave can cross the biological tissue with limited losses and deformation. A study on the influence of frequency variation in signal attenuation is carried especially at interfaces air-tissue and tissue-air. The biological tissue which is characterized by its thickness, its relative dielectric constant and its conductivity, is a sample of fat, localized at a distance of the antenna source which guarantees the inexistence of the attenuation at the amplitude of the electromagnetic fields in the air region. The thickness is about 2.8 cm but the conductivity and the dielectric constant vary with the frequency. The sample is localized in the middle of interest area and it is 50 cm away from transceiver antenna which send signal. A comparison is carried while applying double signals, one is a Gaussian monocycle pulse and other is a simple Gaussian pulse plane and uniform which normally tackles the surface of the samples. To calculate the attenuation of the wave in the biological tissue, we must simulate their propagation by solving Maxwell's equation in these environments. Several methods have been used such as Finite Element Method (FEM) and the Finite Differences Time Domain (FDTD) method. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Attenuation, Electromagnetic Wave Propagation, Finite Differences Time Domain (FDTD), Human Body Tissue, Ultra Wide Band (UWB) Wave

Nomenclature Description D E H ∆z ∆t fc

t0

a C0

Electric flux density vector Electric field intensity vector Magnetic field intensity vector Cell size Time step Frequency of work The delay time from the origin of the Gaussian pulse and the Gaussian monocycle pulse The width of pulse The celerity of light

I. Unit of measure [C/m2] [V/m] [A/m] [m] [s] [Hz]

The aim essentially is the determining the characteristics of the UWB radar for the detection of the beatings of the heart. The search on the biomedical applications of UWB radar will be aimed at the identification of the new devices made possible by the technology in the design, in the development of these devices and in the clinical test of the obtained systems [1],[2]. In previous works established by M. Bilich and D. M. Sullivan [3], [4], [5], [6], the attenuation estimation was done by using the Friis formula [7]. In this one we limit our study to a single biological layer to concentrate the study on the effect of frequency on the signal deformation, especially at the interfaces level Air-Fat and Fat-Air. For this, we calculate the attenuation undergone by a Gaussians pulse in the hand and in the other hand by monocycle pulse during its propagation in a biological tissue at different frequencies and it comparing to determine the suitable frequency at it the wave can cross the sample with the smallest deformation and attenuation of the signal [1], [3]. We use FDTD (Finite Difference Time Domain) who is a technique of modeling general purpose used to

[s] [s] [m/s]

Greek notations µ0 ε0 εr (ω)

Free space permeability [H/m] Free space permittivity [H/m] Permittivity dependant complex relative dielectric constant.

Subscripts k n

Introduction

Distance counter Time counter

Manuscript received and revised October 2011, accepted November 2011

874


M. Chaoui, M. Ketata, M. Lahiani, H. Ghariani

resolve Maxwell's equations. This is useful in the applications where the frequencies of echo are not exactly known, when never a result to wide band is wished. The FDTD tends to provide windows animated by the movement of the electromagnetic field calculated by the model. This type of display is useful to understand what takes place in the model and to help to make sure that the model works correctly [4], [5], [6]. To calculate the attenuation of the electromagnetic wave, the FDTD method is used in the first time to record the electric field amplitude of the wave in each point of space and at each step of time. Then, we deduce the attenuation [8] in biological samples tissue by calculating the signal energy of the transmitted wave in all points of space and by comparing to the incident wave that exists at the interface Air-Fat.The biological tissue is a sample of fat whose thickness is about 2.25 cm, but the dielectric constant and the conductivity vary with frequency. The two sources "antennas" are located at a distance of 4.5 cm from the Air-Fat interface. This distance is lower than 15 cm which is the maximum distance obtained in previous approximations where attenuation of electric field in air area is negligible [1], [2], [3], [5]. This device is shown in Fig. 1. These antennas send a Gaussian monocycle pulse and a Gaussian pulse who can be or not at the same time on a fat sample. The attenuation can be deduced in samples biological tissue which differs by calculating the transmeting signal energy in all points of space and comparing it to the incident one which exists at the interface Air-Fat. The paper is organized as follows. In Section II, onedimensional FDTD method is exposed. Then, Section III highlights the modeling of the source and their propagation environment. Section IV deals of the attenuation of signal energy, while Section V explains the frequency influence on the attenuation. Finally, Section VI summarizes the results and opens to new directions.

II.

The FDTD method is a numerical electromagnetic modelling [4] using a spatial and temporal discretization of Maxwell's equations [5], [6], [7]. These equations can be given as follows relations: ∂D ( t,z ) ∂t

∂H ( t,z ) ∂t

= ∇ ⋅ H ( t,z )

=−

1

µ0

∇ ⋅ E ( t,z )

D ( ω ) = ε 0ε r (ω ) E (ω )

(1)

(2)

(3)

To simplify the formulation of the FDTD equations, it is better to express the Maxwell’s equations using normalized Gaussian units by substituting [1]: E =

ε0 E µ0

(4)

D =

ε0 D µ0

(5)

The electric field is directed along the x-axis and the magnetic induction along the y-axis, the one-dimensional FDTD Maxwell’s equations are written as follows: ∂Dx 1 ∂Hy =− ∂t ε 0 µ0 ∂z ∂Hy 1 ∂Ex =− ∂t ε 0 µ0 ∂z

(6)

The FDTD method is a discritization of Maxwell’s equations in the time and space by using interleaved one dimensional FDTD grid [5], [9]: showing the interdependence between the Ex, Dx and Hy fields; k is the distance counter such that the total simulated distance is z = k·∆z, and n is the time counter such that the total simulated time is t=n·∆t; with ∆z and ∆t stand for the cell size and the time step respectively. Before taking the finite difference approximation of the above expressions, we define an interleaved one-dimensional FDTD grid as showed in Fig. 2. To calculate Hy(k+1/2), for instance, the next values of Ex at k and k+1 are needed. Similarly, to calculate Ex(k+1), the value of Hy at k+1/2 and k+3/2 are needed. Based on that grid, the central difference approximations for both the temporal and spatial derivatives and for reasons of implementation as an iterative algorithm, the Maxwell's equations can be written as:

One-Dimensional FDTD Method

This work uses a one-dimensional FDTD simulation to asses the propagation performance in biological tissues of the Gaussian monocycle pulse or Gaussian pulse proposed for UWB communications.

Fig. 1. Simplified model of propagation Medias


International Review on Computers and Software, Vol. 6, N. 6

875


−

1 2

n−

1

The width in time domain and it is equal a =

( k ) = D x 2 ( k ) +

∆t ⎡ n ⎛ 1⎞ 1 ⎞⎤ ⎛ H y ⎜ k + ⎟ − H yn ⎜ k − ⎟ ⎥ ⎢ 2⎠ 2 ⎠⎦ µ 0 ε 0 ∆z ⎣ ⎝ ⎝ 1

1⎞ 1⎞ ⎛ ⎛ H yn +1 ⎜ k + ⎟ = H yn ⎜ k + ⎟ + 2⎠ 2⎠ ⎝ ⎝ 1 1 ⎤ n+ 1 ∆t ⎡ n + 2 − ⎢ Ex ( k + 1) − E x 2 ( k ) ⎥ µ0ε 0 ∆z ⎣⎢ ⎦⎥

Exn-1/2

Hyn

Exn+1/2

k-2

k-1

k3/2

k

k1/2

K+1

K+1/2

(7)

(8)

k-1

k

K+1

.

1 0.8

K+5/2

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8

k-2

π 2 fc

1

K+2

K+3/2

1

2 2 kp = e : at the work frequency equal to fC and the a maximum of amplitude is set equal to1. Fig. 3 shows the pulse shape of Gaussian pulse and Gaussian monocycle pulse in time domain for fC=4Ghz.

Amplitude of Electric Field E(V/m)

n+

D x

K+2

-1 -4

Gaussian pulse Gaussian monocycle pulse -3

-2

-1

0

1

2

3

4

t(ns)

Fig. 2. FDTD interleaving grid of the E and H fields in space and time

Fig. 3. The Gaussian pulse and Gaussian monocycle pulse

The frequency spectrum of Gaussian pulse and Gaussian monocycle pulse by applying the Fourier Transforming can be written by equation (11) [10]:

III. Model of the Source and the Propagation Environment In this section, we will detail the setting of the propagation environment which is a sample of fat and the performance of the emitted wave which is a Gaussian pulse on a side and the Gaussian monocycle pulse on other side.

TF ⎡⎣ f ( t ) ⎤⎦ = f (ω ) =

f (ω ) =

The biological tissue can be bombarded with many types of pulse. These can be created in different ways including the Hermit pulse which is based on a polynomial function and the monocycles pulse which is the first derivative of a Gaussian pulse [7]. Our study is limited to the Gaussian and Gaussian monocycle waves. Equations (9) and (10) show successively the mathematical formulas of the Gaussian pulse and Gaussian monocycle pulse in time domain [7]:

f (t ) = k p

f (t ) = k p

−iωt ∫ f ( t ) e dt

(11)

−∞

1 k p a3π π e −2πω 2 ⎛ aω ⎞ ⎟ 2 ⎠

(11a) 2

−⎜ 1 f (ω ) = k p a 3ω π e ⎝ 4

(11b)

The results of these waves are represented by Fig. 4. III.2. Simulation Results The frequency is fixed at 4 GHz. The characteristics of the model at this frequency are shown in Table I [8],[10].

2

⎛t⎞ −⎜ ⎟ ⋅t ⋅e ⎝ a ⎠

2π

+∞

The spectrum of these tow waves is given, respectively, by equations (11a) and (11b):

III.1. Model of Source

⎛t⎞ −⎜ ⎟ e ⎝a⎠

1

(9)

TABLE I CONDUCTIVITY AND DIELECTRIC CONSTANT OF THE FAT LAYER AT 4 GHZ [8] dielectric Thickness Conductivity ∂ [S/m] constant ε r [cm]

2

(10)

2.25


0.96

5.5


876


0

propagation. The direct value calculated by equation (12) is 5.305×10-10s but the value given by the simulation is 6.1706×10-10s. This difference is due to the retard experienced by interference between the reflected and incident waves.

Gaussian pulse Gaussian monocycle pulse

-5

-10

YdB(f)

-15

-20

IV.

Calculate of Attenuation of Signal Energy

-25

In this section, we will calculate the attenuation in the biological tissue for the two types of waves and compare one against the other. The energy of signal x(t) is noted by ES, where x(t) is a function of time t. This energy is calculated for a Gaussian pulse at the source antenna by the relation (13) [8]:

-30

-35

-40

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2 9

Frequency (Hz)

x 10

Fig. 4. Frequency spectrums of Gaussian monocycle pulse and Gaussian pulse

+∞

Es =

Amplitude (dB)

Fat

Air

+∞

2

2

dt

(13)

Es is a pair function then:

Air

Es = 2 K p2

∫

⎛t⎞ −⎜ ⎟ e ⎝a⎠

2

2

dt

(13a)

−∞

gaussain pulse antanna 0.903ns 1.053ns 0.7525ns

100

200

300

400

500

-a-

It is assumed that:

600

(a)

x= Amplitude(dB)

Kp

⎛t⎞ −⎜ ⎟ e ⎝a⎠

−∞

0

n (steeps times)

Fat

Air

0.5 0

Air gaussain monocycle pulse antanna

100

⎛t⎞ −⎜ ⎟ e ⎝a⎠

2

(13b)

This leads to:

0.903ns

1.053ns

1

200

300

400

500

n (steeps number times)

1

2 2 ⎡1 ⎤ Es = 2 K p2 ∫ x dx = 2 K p2 ⎢ x3 ⎥ = K p2 ⎣ 3 ⎦0 3 0

0.8725ns

-0.5 0

∫

dt =

−∞

These simulations simulated by Matlab® are shown in Fig. 5(a) for Gaussian pulse and in Fig. 5(b) for Gaussian Monocycle pulse at 0.75ns, 0.903ns and 1.053ns. 0.6 0.4 0.2 0 -0.2 -0.4 0

∫ x (t )

2

-b- 600

(13c)

(b)

For Gaussian monocycle pulse at the source antenna, the energy is given by:

Figs. 5. The spatial amplitude of electric field, afor Gaussian pulse, b- for Gaussian monocycle pulse in different instances at the frequency 4 GHz

0

Es = 2 K p2

It is necessary that the simulation time is sufficiently large that the wave propagates in all regions of space. We can estimate this time by calculating the speed of propagation in air and in the fat. We know that the electromagnetic wave propagates in air with the celerity of light but in fat the speed is given by equation (12) [4],[10]: v=

2C0

ε rmax

∫

⎛t⎞ −⎜ ⎟ te ⎝ a ⎠

2

2

dt

(14)

2

(14a)

−∞

so, it’s true that: 2

⎛t⎞

2

⎜ ⎟ ⎛t⎞ ⎝2⎠ ⎜a⎟ ≤e ⎝ ⎠

2⎛ ⎛t⎞ ⎞ ⎜ −⎜⎝ a ⎟⎠

(12)

then 2

⎛t ⎜ a ⎟ ⎜e ⎝ ⎠ ⎜ ⎝

By calculating those speeds in the different regions of space, we can deduce the time required for the wave propagating from the antenna to the space limits of

2

2 2 ⎞ ⎛t⎞ ⎛ ⎛t⎞ ⎜ ⎟ ⎜ −⎜ ⎟ ⎟ 2 a ⎝ ⎠ ⎝ ⎠ ⎟ ≤e ⎜e ⎟ ⎜ ⎠ ⎝

2 ⎞ ⎛t⎞ −⎜ ⎟ ⎟ a ⎝ ⎠ ⎟ =e ⎟ ⎠

This leads to:



877


⎛t⎞ −⎜ ⎟ t ⋅e ⎝a⎠

0

Es =

2 K p2

∫

2

2

⎛t⎞ −⎜ ⎟ t ⋅e ⎝a⎠

0

dt ≤

2 K p2

−∞

∫

attenuation close to the Air-Fat interface in the Air region, although in this area the attenuation is zero.

2

dt (14b)

−∞

TABLE II CONDUCTIVITY AND DIELECTRIC CONSTANT OF THE FAT LAYER FOR DIFFERENT FREQUENCIES [11]

with the same change of variable: 1

Es ≤ 2 K 2p

⎡1

∫ x dx = 2 K p ⎢⎣ 2 x 2

0

1

Frequency [GHz]

Conductivity ∂ [S/m]

dielectric constant ε

1 3 5 7 9

0.053502 0.13004 0.24222 0.37353 0.13004

0.053502 0.13004 0.24222 0.37353 0.5138

2⎤

2 ⎥ = Kp ⎦0

(14c)

for both sources:

Frequency [GHz]

r

5.447 5.2239 5.0291 4.8476 4.6804

1.2

Es = α K 2p

(14d)

Propagation direction gaussian monocycle pulse

Propagation direction gaussian pulse

1

0.8

In discrete case, the formula becomes:

where:

N

∑ x ( t = n ∆t ) n =1

α K 2p

( t = n ∆t )

2

= α K p2 ( z = k ∆z )

Electric Field E (V/m)

Es =

Transmitted pulse

(14e)

is the maximum amplitude of the

0.4

0.2

Gaussian monocycle pulse

Gaussain pulse antanna

0

-0.2

Transmitted pulse

pulse in each point reported by z = k ∆z . In the air region the attenuation is zero, because we have adopted it to the free space characteristics. The attenuation of the transmitted wave corresponds to the logarithmic ratio of energy at any point in space and the energy of the incident wave which exists on the Air-Fat interface, this attenuation is given by equation (15):

-0.4

Air -0.6

0

0.01

0.02

0.03

0.04

Fat 0.05

0.06

Air 0.07

0.1

Gaussian monocycle pulse antanna

5

0

(15)

where: - Es(k): the transmitted signal energy in kth cells of space. - Eincident: the maximum of signal energy located at the Air-Fat interface [9]. The transmitted and incident waves are shown in (Fig. 6).

V.

0.09

Fig. 6. The transmitted and incident waves vs. distance z

Magnitude (dB)

⎞ ⎟⎟ ⎠

0.08

Z(m)

Gaussian pulse antanna

⎛ E ( n,k ) Att = 10 ⋅ log10 ⎜ s ⎜ E ⎝ incident

Incidente pulse

Incidente pulse

0.6

-5

f= 1GHz f= 3GHz

f= 1GHz

f= 5GHz

f= 5GHz

f= 7GHz

f= 7GHz

f= 9GHz

f= 9GHz

f= 3GHz

-10

-15

-20

-25

Fat

Air 0

0.01

0.02

0.03

0.04

0.05

0.06

Air 0.07

0.08

0.09

0.1

Z (m)

Fig. 7. Attenuation of the signal energy vs. distance z for various frequency

Frequency Influence on the Signal Energy Attenuation

The same remark is valid for the conceited interface Fat-Air, in the Fate region. Indeed in this region the attenuation is normally linear, this is proved in other publications as that of Mr. Staderini [1] or Carlos Bilich [3], [8], but there is a fluctuation in linearity of the attenuation close to Fat-Air interface. This variation is very clear, in the low frequencies noting that at 1 GHz, but the linearity of the attenuation is not remarkable any more: the fluctuation is very small close to 10 GHz.The main observation that can be deduced is that the two curves converge to the same final value, so the two waves undergo the same attenuation. The attenuation in the biological tissue is linear but, at the interfaces Air-Fat, we observe deformations that are

This section studies the influence of frequency variation on the attenuation in the biological tissue. We have to redo the same work at different frequencies, the characteristics of model at those frequencies are shown in Table II. The results are given by Fig. 7. We have shown that both forms of waves have the same way of attenuation. Indeed, we notice for both forms of waves, that the attenuation rises with the increase of frequency. In the low frequencies, it is about 3 dB at 1 GHz, but its value is near 19 dB at 9 GHz. In the other hand, we observe that there is a variation of the Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


878


a( f ) = α ⋅ f 2 + β ⋅ f +γ

more proposed for the Gaussian monocycles wave that the Gaussian wave.

Interpolation of the Attenuation for Determination the Fat Parameters

-100 -200 -300

This section presents an interpolation of the energy attenuation. This interpolation is estimated of first order according to distance Z in the fat layer. The equations of interpolation are given versus the frequency, consequently the coefficients are given according to the latter.By supposing that the fluctuation is of no importance on the level of the interfaces, the fluctuation will be given by: Att ( f ,x ) = a ( f ) ⋅ x + b ( f )

-400

C o e fficie n t "a "

VI.

(17)

b( f ) = λ ⋅ f 2 +φ ⋅ f +ϕ

-500 -600 -700 -800 -900 -1000

3

4

5

6

7

8

9

10 9

frequency (Hz)

x 10

(a) 40

35

(16)

30

coefficient "b"

The results of the interpolation and the attenuation of the signal are given simultaneously on the Fig. 8. 5

25

20

15

0

Magnitude (dB)

10 -5

5 -10

0

3

4

5

6

7

8

frequency (Hz)

-15

9

10 9

x 10

(b) -20

Air -25

0

0.01

0.02

0.03

Fat 0.04

0.05

0.06

Figs. 9. Polynomial coefficients of interpolation “(a)” and ”( b)” vs. frequency

Air 0.07

0.08

0.09

0.1

Z(m)

By identification obtained also after interpolation, the coefficients of dependence are given by the following values: α= -105.6852 β=182.1840 λ= 0.1079 φ=3.3187 ϕ= -7.2140

Fig. 8. Attenuation of the signal energy vs. distance z for various frequency (1GHz-9GHz)

The polynomial coefficients of interpolation for various frequencies are represented in the Table III TABLE III CHARACTERISTICS VALUES OF INTERPOLATION OF ATTENUATION ON THE FAT LAYER FOR DIFFERENT FREQUENCIES f(GHz) a(f) b(f) 3 -187.1805 4.2637 4 -306.7421 10.0518 5 -350.5855 11.8868 6 -457.3836 16.9949 7 -548.8755 21.1491 8 -655.5482 26.1381 9 -767.1507 31.3707 10 -883.1874 36.8247

The attenuation can be written by taking account of relations (17):

( + (λ ⋅ f

) +φ ⋅ f +ϕ)

Att ( f ,x ) = α ⋅ f 2 + β ⋅ f + γ ⋅ x + 2

(18)

The energy transmitted was deduced in function to incidental energy:

To see the dependence of the coefficients a and b with the frequency, the following Figs. 9 show the layout of these coefficients versus the frequency. It is clear that the two characteristic values are dependent in a linear way of the frequency. Indeed, the characteristic value "a" decreases relatively when the frequency increases and conversely for "b". Thereafter we expressed the characteristic values according to the frequency by the following relations:

Et ( f ,x ) = Att ( f ,x ) ⎞ ⎛ = exp ⎜ Ln (10 ) ⋅ ⎟ ⋅ Eincident ( f ) 10 ⎝ ⎠

(19)

by replacing equation (18) in (19) this gives:



879


Et ( f ,x ) =

(

)

⎛ ⎡ α ⋅ f 2 + β ⋅ f + γ ⋅ x +⎤ ⎞ ⎜ ⎢ ⎥⎟ ⎜ ⎢+ λ ⋅ f 2 + φ ⋅ f + ϕ ⎥⎟ ⎜ ⎢ ⎣ ⎦⎥ ⎟ ⋅ = exp ⎜ Ln (10 ) ⎟ 10 ⎝ ⎠ ⋅ Eincident ( f )

(

)

(

⎛ λ ⋅ f 2 +φ ⋅ f +ϕ ⎜ Et ( f ,x ) = exp Ln (10 ) ⎜ 10 ⎝

(

equations which illustrate the laws of the UWB wave's propagation. Then we explained the FDTD method that we adopted it in our problem and we studied the attenuation of the signal energy and the influence of frequency on the signal deformation. In spite of the simplicity of the propagation model area (Air and Fat), we succeeded to have a satisfactory results for the comparison of the performances of a Gaussian pulse and its first derivative (Gaussian monocycle pulse) during their propagation in the biological tissue. For the two form of waves in the low frequencies, the deformation at Air-Fat and Fat-Air interfaces is very important. The attenuation is low, but it is not linear, the two curves converge almost to the same values (about -3dB at 1GHz and at 10GHz it is above 22dB) In the high frequencies, the deformation is less important but it is more visible for the Gaussian pulse monocycle than for the Gaussian pulse. We also notice that the linearity of the attenuation is clearer in both forms of waves (almost with the same slope). At the level of the interfaces, the deformation of the signal is almost symmetric for Gaussian monocycle but this remark is not valid for Gaussian pulse. In this simulation, we adapt ourselves a very simple relief which is identified by the dielectric constant, the width of the sample and the electric conductivity of the fat. Our approach of research makes it possible to give the characteristics of the biological sample without referring to the data, rather containing following the signal propagated in the sample. These parameters are real constants, we could take the other approaches to describe this model such as for Cole-Cole, where the parameters become complex. Consequently, we can start the program from the Z transformer, which can be a future work. We can also remain the modeling in 3D for the study of the scattering characteristics of the heart.

(19a)

) ⎞⎟ ⋅ )

⎟ ⎠

⎛ ⎞ α ⋅ f 2 + β ⋅ f +γ ⋅ exp ⎜ Ln (10 ) ⋅ x⎟⋅ ⎜ ⎟ 10 ⎝ ⎠ ⋅Eincident ( f )

(19b)

This formula is according to M.Bilich [10] and M. Staderini[1]: ⎛η 2⎞ Et ( f ,x ) = ⎜ 1 Γ ⎟ exp ( −2 ⋅ α 2 ⋅ x ) ⋅ Eincident ( f ) ⎝ η2 ⎠

(20)

By identification equations (19b) and (20):

(

⎛ λ ⋅ f 2 +φ ⋅ f +ϕ ⎛ η1 2 ⎞ ⎜ 10 exp Ln Γ = ⋅ ( ) ⎜ ⎟ ⎜ 10 ⎝ η2 ⎠ ⎝

) ⎞⎟ ⎟ ⎠

(21)

and:

α 2 = − Ln (10 ) ⋅

(α ⋅ f

2

+ β ⋅ f +γ

)

(22)

20

The equations (21) and (22) make it possible to deduce:

References [1]

⎛η 2⎞ α 2 = 20.087 and ⎜ 1 T ⎟ = 0.669 ⎝ η2 ⎠

[2]

On the other hand Bilich found: ⎛ η1

α 2 = 27.17 and ⎜

⎝ η2

[3]

2⎞ T ⎟ = 0.707 ⎠

[4]

A light difference is distinguished between those deduced and those given by Bilich, this is due mainly to the order of interpolation in α (which is taken equal to 2) and in addition to the negligence of the parameter J which is considered small in front of ε (equation (3)).

VII.

[5]

[6]

[7]

Conclusion

[8]

In this work, we began by presenting Maxwell's


M. Staderini, UWB Radars in Medicine, IEE AESS Systems Magazine, volume 1, no. 3, Mars 2000,Pages 1-10. F. Thiel, M. Hein, U. Schwarz, J. Sachs, and F. Seifert, Combining magnetic resonance imaging and ultra wide band radar: A new concept for multimodal biomedical imaging, Review of Scientific Instruments, 2009, 80 014302 Bilich and C. G, Feasibility of Dual UWB Heart Rate Sensing and Communications Under FCC Power Restrictions, Third International Conference on Wireless and Mobile Communications, ICWMC (2007) Dennis M. Sullivan, Electromagnetic simulation using the FDTD method (IEEE Press Series on RF and Microwave Technology, Series Editors, 2000.) Shuichi Aono, Masaki Unno and Hideki Asai, AlternatingDirection Explicit FDTD Method for Three-Dimensional FullWave Simulation, Electronic Components and Technology Conference, 2010. Dennis M. Sullivan, A frequency-dependent FDTD method for biological applications, IEEE transactions on microwave theory and techniques, vol. 40, no. 3, march 1992. PulsON® Technology Overview, Huntsville, AL, USA (2001). Huntsville, AL, USA; July 2001. Bilich, C. G, “UWB Radars for Bio-Medical Sensing: Attenuation Model for Wave Propagation in the Body at 4GHz”, Informatica e


880


Telecomunicazioni, Technical Report DIT-06-051University of Trento, 2006. [9] August W. Rihaczek, Signal energy Distribution in Time and Frequency, IEEE Transactions on information Theory, vol. IT-14, no. 3, May 1968. [10] Bilich, C, A Simple FDTD Model to Asses the Feasibility of Heart Beat Detection using commercial UWB communication devices, Technical Report DIT-07-033 38050. [11] Camelia Gabriel , Dielectric properties of biological tissue: Variation with age, Wiley-Liss, Vol 26, n 7, pages S12–S18, 2005

Authors’ information University of Sfax, Tunisia National Engineering School of Sfax, Laboratory of Electronics and Information Technologies (LETI) E-mail: [email protected] Mondher Chaoui was born in Sfax, Tunisia, in May 1968. He received the Engineering Degree in Electrical Engineering, the DEA in Electronics from the National Engineering School of Sfax in 1994, 1998, respectively and the Doctorate of Engineer in Electronics from the same School in 2006. In 2003 he joined the Sfax University for teaching. His current research interests are design of circuits in medical electronic, mutual inductance link in the implantable systems, RFpower amplifiers in medical electronic and the techniques Radar for measurement of the physiological parameters of the human body. Moez Ketata was born in Sfax, Tunisia, on August 1977. He received the Electrical Engineering Degree from the“National Engineering School of Sfax -Tunisia” in 2004, the “MASTER” degree on electronics in 2007. He is currently working toward a “Doctorate on Electrical engineering” at the same university. Since 2007, he has worked as contractual assistant at the Faculty of Science of Gafsa. Also he is member in the “LETI” Laboratory ENIS Sfax. His main research interests are the techniques Radar for measurement of the physiological parameters of the human body. Mongi Lahiani was born in Sfax, Tunisia, in 1957. He received the Electrical Engineering Degree from the University of Sciences and Techniques of Sfax-Tunisia in 1984 and the Doctorate of Engineer in Measurement and Instrumentation from University of Bordeaux1, France, in 1986. Since 1988, he has been a Professor at Sfax University. He joined the Sfax National School of Engineering in 1990 and he is a Professor in the fields of analog electronics and microelectronics. His research interests are design of circuits in medical electronic and integration of microwave circuits of microstrips with screen-printed thick films technology. Hamadi Ghariani was born in Sfax, Tunisia, in July 1956. He received the Electrical Engineering Degree from the “University of Sciences and Techniques of Sfax-Tunisia” in 1981, the “DEA” degree in 1981 and his “Doctorate of engineer” in 1983 in “Measurement and Instrumentation” from the “University of Bordeaux France”. He joined the National Engineering School of Sfax since 1984. Actually he is a Professor in the same School. His research activities have been devoted to several topics: Medical Electronic; Communication systems for Medical Telemetry, Measure and Instrumentation.



881


Recommendation Systems Through Semantic Web Usage Mining Mehrdad Jalali1, Norwati Mustapha2 Abstract – Internet users are drowned in all kind of available information. However, only a tiny part of it is usually relevant to their preferences. Web usage mining which extracts knowledge for usage and clickstream data has become the subject of exhaustive research, as its potential for Web-based personalized services, prediction of user near future intentions, adaptive Web sites, and customer profiling are recognized. Moreover, semantic web aims to enrich the WWW by machine processable information which supports the user in his tasks. Semantic clickstream mining which integrates semantic in Web usage mining processes aims to improve the quality of the Web usage mining systems. Given the primarily syntactical nature of data Web usage mining operates on, the discovery of meaning is impossible based on these data only. Therefore, formalizations of the semantics of Web resources and navigation behavior are increasingly being used. In this paper, we discuss the interplay of the Semantic Web with Web usage mining and also we give an overview of where the two research areas meet today. Moreover a proposed framework to integrate semantic Web and Web usage mining discuss in the rest of the paper. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Web Usage Mining, Clickstream Mining, Semantic Web, Semantic Web Usage Mining

I.

will be illustrated in the end of this section.

Introduction

Semantic Web Usage Mining (SWUM) or Semantic Clickstream Mining aims to integrate two research areas Semantic Web and Web Usage Mining for obtaining more fine-grained and meaningful user behaviours in the Web environment. To better understand user's next intentions from observing him while navigating on a Website, all semantically interaction data needs to be tracked as well as tracking clickstream data. The most important issue facing in the classical Web Usage Mining system is quality of the results. The aim of this paper is to give an overview of where the two areas meet today, and what we can do to improve the results of integrating semantic Web and Web Usage Mining. The remainder of this paper is organized as follows: Section 2 covers a brief overview of the areas Semantic Web and Web Usage Mining. Section 3 describes some related research about semantic Web usage mining and introduces proposed framework for integrating semantic Web and Web usage mining. Finally, section 4 concludes the current study and sheds light on some directions in the future works.

II.

II.1.

Web Usage(Clickstream) Mining

In general, Web mining can be characterized as the application of data mining to the content, structure, and usage of Web resources [1], [2]. The goal of Web mining is to automatically discover local as well as global models and patterns within and between Web pages or other Web resources. However, Web mining tools aim to extract knowledge from the Web, rather than retrieving information. Research on Web mining is classified into three categories, which are Web structure mining that identifies authoritative Web pages, Web content mining that classifies Web documents automatically or constructs a multilayered Web information base, and Web usage mining that discover user access patterns in navigating Web pages [3]. The goal of Web usage mining, in particular, is to capture and model Web user behavioral patterns. The discovery of such patterns from the enormous amount of data generated by Web and application servers has found a number of important applications. Among these applications are systems to evaluate the effectiveness of a site in meeting user expectations [4], techniques for dynamic load balancing and optimization of Web servers for better and more efficient user access [5], and applications for dynamically restructuring or customizing a site based on users’ predicted needs and interests. From the data-source perspective, both Web structure and Web content mining target the Web content, while

The Web Usage Mining and Semantic Web

In the first part of this section, we cover some backgrounds in the WUM systems. In the second part, we recall our understanding of semantic Web. A brief discussion about integrating these two areas


882


Mehrdad Jalali, Norwati Mustapha

A) Domain ontology acquisition: The process of acquiring, maintaining and enriching the domain ontologies is referred to as “ontology engineering”. For small Web sites with only static Web pages, it is feasible to construct a domain knowledge base manually or semi-manually. The outcome of this phase is a set of formally defined domain ontologies that precisely represent the Web site. Good representation should provide machine understandability, the power of reasoning, and computation efficiency. B) Knowledge base construction: While the first phase generates the formal representation of concepts and relations among them, the second phase, knowledge base construction, can be viewed as building mappings between concepts or relations on the one hand, and objects on the Web. The goal of this phase is to find the instances of the concepts and relations from the Web site’s domain, so that they can be exploited to perform further data mining tasks. Information extraction methods play an important role in this phase. C) Knowledge-enhanced pattern discovery: Domain knowledge enables analysts to perform more powerful Web data mining tasks. For example, semantic knowledge may help in interpreting, analyzing, and reasoning about usage patterns discovered in the mining phase. In the following we introduce semantic Web which can be integrated to the Web usage mining process.

Web usage mining targets the Web access logs. Web usage mining (WUM) comprises three major processes: data pretreatment, data mining, and pattern analysis [3]. Pretreatment performs a series of processing on Web log files, which are data conversion, data cleaning, user identification, session identification, path completion, and transaction identification. Next, mining algorithms are applied to extract user navigation patterns. A navigation pattern represents the relationships among Web pages in a particular Web site. Some pattern analyzing algorithm is applied to extract data from data mining part for the recommendation system. Recently, a number of Web usage mining (WUM) systems have been proposed to predict user’s preferences and their navigation behaviors. More recently, Web usage mining techniques have been proposed as another user-based approach to personalization which alleviates some of the problems associated with collaborative filtering. In particular, Web usage mining has been used to improve the scalability of personalization systems based on traditional CF-based techniques. In [6] we advance an architecture for online predicting in Web usage mining recommendation system and propose a novel approach to classifying user navigation patterns for predicting users’ future requests. The approach is based on using the longest common subsequence (LCS) algorithm in classification part of the system. All of these works attempt to find architecture and algorithm to improve accuracy of personalized recommendation, but the accuracy still does not meet satisfaction especially in large-scale Websites. Fig. 1 illustrates the state of the proposed system.

II.2.

Semantic Web

The Semantic Web aims to obtain machineunderstandable information from WWW which is based on a vision of Tim Berners-Lee, the inventor of the WWW. The great success of the current WWW leads to a new challenge: a huge amount of data is interpretable by humans only; machine support is limited. He suggested to enrich the Web by machine-processable information which supports the user in his tasks. For instance, today’s search engines are already quite powerful, but still frequently return overly large or inadequate lists of hits. Machine-processable information can point the search engine to the relevant pages and improve the quality of the results. Fig. 2 shows the layers of the Semantic Web as suggested by Berners-Lee. This architecture is discussed in detail for instance in [8] and [9].

Fig. 1. A WUM system framework [6]

To improve the quality of the results of the WUM systems, domain knowledge about a Web site can be integrated to the WUM process. Domain knowledge can be integrated into the Web usage mining process in many ways. This includes leveraging explicit domain ontologies or implicit domain semantics extracted from the content or the structure of documents or Web site. In general, however, this process may involve one or more of three critical activities [7]:

Fig. 2. The layers of the Semantic Web



883


personalization process based on Web usage mining is given in the following. Our main research question is; Can usage patterns reveal further relations to help build the Semantic Web? This field is still rather new, so we will only describe an illustrative selection of research approaches. Ypma and Heskes propose a method for learning content categories from usage [10]. They model navigation in terms of hidden Markov models, with the hidden states being page categories, and the observed request events being instances of them. Their main aim is to show that a meaningful page categorization may be learned simultaneously with the user labeling and intercategory transitions; semantic labels (such as “sports pages”) must be assigned to a state manually. The resulting taxonomy and page classification can be used as a conceptual model for the site, or used to improve an existing conceptual model. Chi et al.[7] identify frequent paths through a site. Based on the keywords extracted from the pages along the path, they compute the likely “information scent” followed, i.e. the intended goal of the path. The information scent is a set of weighted keywords, which can be inspected and labeled more concisely by using an interactive tool. Thus, usage creates a set of information goals users expect the site to satisfy. These goals may be used to modify or extend the content categories shown to the users, employed to structure the site’s information architecture, or employed in the site’s conceptual model. Stojanovic, Maedche, Motik, and Stojanovic [11] propose to measure user interest in a site’s concepts by the frequency of accesses to pages that deal with these concepts. They use these data for ontology evolution: Extending the site’s coverage of high-interest concepts, and deleting low-interest concepts, or merging them with others. The combination of implicit user input (usage) and explicit user input (search engine queries) can contribute further to conceptual structure. User navigation has been employed to infer topical relatedness, i.e. the relatedness of a set of pages to a topic as given by the terms of a query to a search engine. A classification of pages into “satisfying the user defined predicate” and “not satisfying the predicate” is thus learned from usage, structure, and content information. An obvious application is to mine user navigation to improve search engine ranking [12]. Many approaches use a combination of content and usage mining to generate recommendations. For example, in contentbased collaborative filtering, textual categorization of documents is used for generating pseudo-rankings for every userdocument pair [8]. In [9], ontologies, IE techniques for analyzing single pages, and a user’s search history together serve to generate recommendations for query improvement in a search engine. In [13], Authors have presented a general framework for using domain ontologies to automatically characterize usage profiles containing a set of structured Web objects. Their motivation has been to use this framework in the

Today, it is almost impossible to retrieve information with a keyword search when the information is spread over several pages. Consider the query for Web mining experts in a company intranet, where the only explicit information stored are the relationships between people and the courses they attended on one hand, and between courses and the topics they cover on the other hand. In that case, the use of a rule stating that people who attended a course which was about a certain topic have knowledge about that topic might improve the results. The process of building the Semantic Web is today still under way. Its structure has to be defined, and this structure should be brought to life. Given the primarily syntactical nature of data Web usage mining operates on, the discovery of meaning is impossible based on these data only. Therefore, semantic knowledge of Web documents and navigation behaviors can be utilized in recommendation systems for predicting different types of complex Web document and object based on underlying properties and attributes especially in large-scale and dynamic Web sites. Mapping between user navigation transactions in Web usage mining to semantic transaction based on concepts and objects can improve the accuracy of the Web usage mining personalization system.

III. Semantic Web Usage Mining Our goal in this section is to provide a road map for the integration of semantic and ontological knowledge into the process of Web usage mining. Semantics can be utilized for Web Usage Mining for different purposes which are introduced in this section. To better integrate semantic Web and WUM it would be desirable to have a rich semantic model of content and structure of a site. This model should capture the complexity of the manifold relationships between the concepts covered in a site, and should be ”built into” the site in the sense that the pages requested by visitors are directly associated with the concepts and relations treated by it. This leads to semantic Web usage mining. Semantic Web usage mining involves the integration of domain knowledge into Web usage mining [7]. Utilizing semantic knowledge can lead to deeper interaction of the Website’s user with the site. Integration of domain knowledge allows such systems to infer additional useful recommendations for users based on more fine grained characteristics of the objects being recommended, and provides the capability to explain and reason about user actions. The interpreting, analysing, and reasoning about usage patterns discovered in the mining phase can be done by using semantic knowledge. Moreover, it can improve the quality of the recommendations in the usage-based system. Several studies have considered various approaches to integrate content-based semantic knowledge into traditional usage-based recommender systems. An overview of the existing approaches as well as a some framework for integrating domain ontologies with the



884


sessions to ground any guesses about relevancy. To enhance the discovered patterns’ quality, the authors in [15] propose using metadata about the content that they assume is stored in a domain ontology. Their approach comprises a dedicated pattern space built on top of the ontology, navigation primitives, mining methods, and recommendation techniques. Web usage mining (WUM) approaches often use terms and frequencies to represent a Web site for the mining process. In [16], the authors show that these representations lead to poor results. Therefore, it is proposed to perform a semantic Web usage mining process to enhance quality of the mining results. In this paper it was used a concept-based Web usage mining process to generate more semantically related results. The approach was used to enhance a real Web site and it was evaluated by comparing it with four different WUM methods. It was defined two quality measures (interest and utility) in order to evaluate the results. These measures are obtained using surveys to 100 visitors of the site. Based on interest and correlation measures, it was proved that concept-based approach allows obtaining results closer to visitors’ real browsing preferences. Moreover, information produced by the proposed approach lead to the discovery of enhancements. The proposed method also finished the generalization task in a few minutes which is not too much compare with other methods. In [10], they proposed the integration of semantic information drawn from a Web application’s domain knowledge into all phases of the Web usage mining process (preprocessing, pattern discovery, and recommendation/prediction). The goal is to have an intelligent semantics-aware Web usage mining framework. This is accomplished by using semantic information in the sequential pattern mining algorithm to prune the search space and partially relieve the algorithm from support counting. In addition, semantic information is used in the prediction phase with low order Markov models, for less space complexity and accurate prediction that will help solve ambiguous predictions problem. Experimental results show that semantics-aware sequential pattern mining algorithms can perform 4 times faster than regular non-semantics-aware algorithms with only 26% of the memory requirement. Fig. 5 illustrates the proposed architecture for a semantic Web usage mining system which can be used in a recommender system. In the offline phase of the system to perform semantic data pretreatment, Web site ontology and a knowledgebase which are created based on the content and structure of the Web site can be utilized in the process of this module. On the other hand to create semantic usage data which further will be used in semantic navigation pattern, and to understand semantic knowledge about user semantically’ sessions in a particular website, this module needs to integrate with those ontology and knowledgebase. In the next module, semantic navigation patterns will be extracted from the sessions by utilizing semantic

context of Web personalization, going beyond page-level or item-level constructs, and using the full semantic power of the underlying ontology. They considered a Web site as a collection of objects belonging to certain classes (resulting in a concept hierarchy of Genre’s a portion of which is shown in Fig. 3). Given a collection of similar user sessions (e.g., obtained through clustering) each containing a set of objects, they have shown how to create an aggregate representation of for the whole collection based on the attributes of each object as defined in the domain ontology (Fig. 4). This aggregate representation is a set of pseudo objects each characterizing objects of different types commonly occurring across the user sessions. They have also presented a framework for Web personalization based on domain-level aggregate profiles.

Fig. 3. The Ontology for a movie Web site

Fig. 4. An Example of an Object in Class Movie

In [14], they proposed an approach to track user interaction data and preserving semantic knowledge on complex and interactive Web sites. They showed that the approach comes with some major enhancements compared to some existing solutions. The usage of Microformats enables an easy integration into existing Web sites and allows then to interrelate data on these sites (Microformats are small patterns of HTML to represent commonly published things like people, events, blog posts, reviews and tags in web pages. Microformats enable the publishing of higher fidelity information on the Web; the fastest and simplest way to provide feeds and APIs for the information in your website.). This also allows them to obtain fine-grained information connected with semantic knowledge that opens new chances to personalize Web sites. Recommender systems rely on relevance scores for individual content items; in particular, pattern-based recommendation exploits co-occurrences of items in user Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


885


[6]

clustering algorithm. In the online phase, the system recommends some pages and concepts which the current users intend to navigate them through the particular Web site. This phase in similar to our work which is described in [6]. In summary, all of these works attempt to find reference architecture and framework to improve quality of the Web usage mining systems by integrating semantic web to WUM, In the proposed framework, we advance a framework to integrate semantic web and WUM, which by using appropriate semantic clustering algorithm and well-done ontology and knowledgebase design, the quality of the semantic Web usage mining can be enhanced as future work.

[7]

[8]

[9]

[10]

[11]

[12]

[13] [14]

[15]

[16]

M. Jalali, et al., "WebPUM: A Web-based recommendation system to predict user future movements," Expert Systems with Applications, vol. 37, pp. 6201-6212, 2010. E. H. Chi, et al., "Using information scent to model user information needs and actions and the Web," in Proceedings of the SIGCHI, 2001, pp. 490-497. P. Melville, et al., "Content-boosted collaborative filtering for improved recommendations," in Menlo Park, CA; Cambridge, MA; London; AAAI Press; MIT Press; 1999, 2002, pp. 187-192. S. Parent, et al., "An adaptive agent for web exploration based on concept hierarchies," in In Proceedings of the 9th International Conference on Human Computer Interaction, 2001. N. R. Mabroukeh and C. I. Ezeife, "Using domain ontology for semantic web usage mining and next page prediction," in CIKM '09 Proceeding of the 18th ACM conference on Information and knowledge management, 2009, pp. 1677-1680. L. Stojanovic, et al., "User-driven ontology evolution management," Knowledge Engineering and Knowledge Management: Ontologies and the Semantic Web, pp. 133-140, 2002. C. Kemp and K. Ramamohanarao, "Long-term learning for web search engines," Principles of Data Mining and Knowledge Discovery, pp. 243-311, 2002. H. Dai and B. Mobasher, "Using ontologies to discover domainlevel web usage profiles," Semantic Web Mining, p. 35, 2002. T. Plumbaum, et al., "Semantic web usage mining: Using semantics to understand user intentions," User Modeling, Adaptation, and Personalization, pp. 391-396, 2009. M. Adda, et al., "Toward recommendation based on ontologypowered web-usage mining," IEEE Internet Computing, pp. 4552, 2007. S. R os and J. Velásquez, "Semantic web usage mining by a concept-based approach for off-line web site enhancements," 2008 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, pp. 234-241, 2008.

Fig. 5. The proposed framework

Authors’ information IV.

1

Conclusion and Future Work

In this paper, we have studied the integration of the two fast developing research areas Semantic Web and Web usage mining. We discussed how Semantic Web usage mining can improve the results of Web usage Mining systems by exploiting the semantic structures in the process of the Web usage mining. Moreover, a proposed framework to integrate semantic Web and Web usage mining discussed in this paper. As a future direction, we plan to develop a semantic Web usage mining system which uses the proposed framework.

Department of Computer Engineering, Mashhad Branch, Islamic Azad University, Mashhad, Iran. E-mail: [email protected] 2 Department of Computer Science, Universiti Putra Malaysia, Malaysia. E-mail: [email protected]

References [1]

[2]

[3]

[4] [5]

R. Cooley, et al., "Web mining: information and pattern discovery on the World Wide Web," in Ninth IEEE International Conference on Tools with Artificial Intelligence, Newport Beach, CA, USA, 1997, pp. 558-567. J. Srivastava, et al., "Web usage mining: discovery and applications of usage patterns from Web data," ACM SIGKDD Explorations Newsletter, vol. 1, pp. 12-23, 2000. B. Mobasher, et al., "Creating adaptive Web sites through usagebased clustering of URLs," in Knowledge and Data Engineering Exchange, Chicago, IL, USA, 1999, pp. 19-25. M. Spiliopoulou, "Web usage mining for web site evaluation," Communications of the ACM, vol. 43, pp. 127-134, 2000. J. E. Pitkow, et al., WebVis: A Tool for World Wide Web Access Log Analysis: Graphics, Visualization & Usability Center, Georgia Institute of Technology, 1994.



886


Prototyping Mobile Graphical Authentication Martin Mihajlov1, Borka Jerman Blažič 2, Dragan Tevdovski3

Abstract – As users can only make informed choices when the proposals being discussed are meaningful to them, enabling users to envision and make sense of those proposals is an essential element of all approaches to system design. The focus of this study is the conceptualization of a graphical authentication mechanism appropriate for ubiquitous environments. Therefore, the experiment presented in this paper uses the paper prototyping method is used to effectively simulate graphical authentication and distinguish the user preference between two system approaches in ubiquitous environments. Complementary, both a high fidelity paper and a mobile prototype is used in order to evaluate popular graphical authentication concepts which are based on separate cognitive functions: recall and recognition. This experiment uses a between group design where each participant evaluates a prototype for both authentication concepts in the same medium. The results of the study demonstrate that recognition-based graphical authentication mechanisms are more suitable for ubiquitous environments than recall-based systems. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Graphical Authentication, Mobile Prototyping, System Evaluation

I.

The variance in fidelity between a prototype and the final product can be along several dimensions. They include: • breadth of features, • degree of functionality, • similarity of interaction, and • aesthetic refinement. In [9] the authors note that a prototype that compromises on one or more of these four dimensions in a way that is obvious to the user is a low-fidelity prototype. Hence, as described in [10], a low fidelity prototype is simple and built with the intent of visualizing design ideas at very early stages of the design process. Comparatively, high-fidelity prototypes are representations composed of predefined components and scripted interactions. Thus, without compromising the noted dimensional aspects and the increased investment needed for development, the users can directly interact with the prototype. Another implication in prototyping is the choice between paper and computer as the prototype medium, which has implications in the realism of the representation, the type of available usability testing methods and the ability of users to participate in the design process [11]. When compared to a digital medium, paper obviously doesn't respond to either mouse or keyboard input, thus the differing mediums support different aspects of the interaction process. In order to demonstrate the interaction process in lowfidelity prototyping tests, the examiner needs to create a pretend-environment that simulates a fully functional system. With a paper medium this includes manipulating sheets of paper in response to the user’s behavior.

Introduction

Paper prototyping is a widely used and validated technique for exploring, communicating, and evaluating early interface designs [1].It is an established approach in designing interactive systems for the development of basic software versions that help users and designers to understand the possible alternatives [2] [3].Conversely, prototypes are used to examine the aesthetics, content and interaction techniques from the perspective of the user. By gathering data on user mistakes and comments, designers and usability professionals can trace usability problems at an early stage [4][5]. In [6] the author states that paper prototyping allows for testing early design ideas at an extremely low cost, which allows for fixing usability problems before implementing something that doesn't work. This implies the many benefits of paper prototyping during the design process, such as [7]: • externalizing design ideas with low investment, • generating and testing numerous alternatives early in the design cycle, • iterating a design many times prior to committing to an implementation, and • focusing evaluation on macro-level issues such as major interface screens and overall interaction. Alternatively, paper prototyping has some inherent limitations, such as the lack of complete realism in the interaction or understanding and revising dynamic behaviors [8]. However, it is generally considered that these limitations are a worthwhile trade-off for the ability to explore numerous alternatives early in the design cycle.


887


M. Mihajlov, B. Jerman-Blažič, D. Tevdovski

Furthermore, paper prototypes allow for quick changes while exploring interactions by sacrificing some realism. On the other hand, computer prototyping requires a better defined interaction flow before user testing as it usually allows for pre-programmed responses to user behavior and remote recording of user interactions. However, high-fidelity computer prototypes may reduce design effectiveness as development tools limit the creative flow and slows down the prototyping process [12]. Low-fidelity computer prototyping require less skill and allow for quick changes by sacrificing the range of interaction techniques available in the prototype. The need and effectiveness of low and high fidelity prototypes have been analyzed in several studies. [9] compared usability problems discovered by low and high fidelity prototypes, and conclude that substantially the same sets of usability problems were found under both conditions. [10] also reached the same conclusion with the added knowledge that subjects prefer computer over paper prototypes. When considering ubiquitous environments, in [13] the authors determine that low fidelity paper prototyping is insufficient for supporting requirements, such as scalability, but a prototype with higher fidelity and automation levels can enhance the quality of interaction data available for evaluation. Snyder [1] describes how to prepare, conduct and interpret the results of paper prototyping. A common technique for evaluating paper prototypes is user evaluation [14], where a user informally works through several controlled tasks and the design team identifies where the prototype met or didn’t meet user expectations. This can be complemented with other evaluation techniques such as cognitive walkthroughs [15] and heuristic evaluations [16]. Most evaluation techniques utilize the Wizard of Oz approach [17], where when a user interacts with the interface, a facilitator changes the representation in the background to simulate an interaction with a fully functional application.

II.

applications deployed in ubiquitous environments have not yet been established. Ubiquitous environments are dynamic and typically complex with an "on-the-go" context of use, hence usability problems are best discovered in situations representative of the real world [18]. Mobile prototyping studies are rarely performed due to the challenges inherent to mobile use [19]. In an onthe-street study [20] the authors used a table PC-sized cardboard box to simulate screens of a mobile user interface. They reported a great difficulty in using this prototype in the field as prototype components were difficult to manage in such an environment. [21] discovered that the fragility of just using cards and paper misled participants about the form factor of the final product and used card-holding wooden frames with dimensions approximating the target device. Extending these studies, a comparison of traditional paper prototype to pseudo-paper prototype used in a mobile contextually relevant lab-based protocol showed that the later allows participants to identify more unique usability problems [22]. Reflecting on authentication mechanisms, there is no research of either paper or mobile prototyping during the development process of such mechanisms. Considering the specificity of ubiquitous environments and the most current research overview of usability and security in graphical authentication [23], it is necessary to test graphical authentication concepts using both a high fidelity paper prototype and a high fidelity mobile prototype in order to encapsulate different aspects of the interaction process during evaluation. The prototyped graphical authentication concepts evaluated in this experiment differ on the mental task necessary to memorize and reproduce the graphical password: recall and recognition. The recall-based graphical authentication concept, codenamed ImagO, has its roots in Blonder-style passwords [24] and is based on Microsoft’s’ approach to single image recognition. In ImagO the user is presented with a large photographic image that contains a sizeable amount of distinct elements. The elements in question represent either artificial objects or living beings, each defined as a separate clickable region. To authenticate, the user clicks on the distinct elements in the image, thus entering the password (Fig. 1).

Mobile Prototyping and Graphical Authentication

Although paper prototyping is a reputable method for evaluating usability in early design iterations, the definite principles of paper prototyping for evaluating

Fig. 1. ImagO concept example



888


The recognition-based graphical authentication concept, codenamed ImagePass, is based on VIP [25]. In ImagePass the user is asked to enter the graphical password by selecting a series of images representing objects from a selection grid which contains both password and decoy images (Fig. 2). If both the sequence and the selected images are correct the user is granted permission to access the system. The improved memorability and the suitability of photographic images for authentication have been demonstrated in [26].

small number of users who necessarily don't need to be chosen carefully [27]. Nevertheless, as this study investigates user behavior in ubiquitous environments the selected participants belong in the generation X and generation Y cohorts. Hence, thirty-eight participants (20 male, 18 female, age range 17-36, average age 27.3) were recruited for this study. All of the participants had normal vision, varied touch-screen gadget experience, and used mobile Internet for professional or leisure purposes on a frequent basis. The paper prototype versions of the graphical authentication mechanism were created with A4 paper. The interactive prototype versions utilized a Samsung Galaxy S touch screen mobile phone with Android 2.2 OS. All experiment sessions were conducted at the Laboratory for Open Systems and Networks at the Jozef Stefan Institute and the E-business Laboratory at the Ss. Cyril and Methodius University. III.3. Experiment Design For the purpose of this experiment, six large images with a substantial number of recognizable distinct elements were prepared for the ImagO prototypes and fifty small single-object images were prepared for the ImagePass prototype. For the paper prototype the images were printed in color on A4 sized paper. Single page color printouts were used for each image for prototyping the ImagO concept while the ImagePass concept used multiple color cut-outs that were managed by the examiner during the session as necessary. The paper prototype was laid out on a table with the examiner performing the role of the computer by moving parts of the interface in accordance with users’ actions. In the mobile prototype a mock-up app was used which showed images with undefined click areas and a pre-programmed tap response regardless of where on the screen the participant decided to press. In order to shift the participants focus to authentication as a secondary task and also to give context to the authentication process, in the initial version of the experiment three different authentication scenarios were developed: email login, forum login and system/phone unlock. However, this approach was discarded as the nature of the scenarios introduced a strong bias in the results. For example, during the email scenario the participants had a strong dislike for the authentication mechanism, as they immediately felt that the authentication mechanism was inappropriate for mail use. This approach could be feasible in future studies when the appropriate environment is determined for specific systems.During the session the participants were asked to think aloud while the examiner occasionally focused the direction to their security concerns. All of the participants’ comments were recorded with a digital voice recorder. The content of the collected comments was analyzed to specifically define the variables to be evaluated.

Fig. 2. ImagePass concept example

III. The Prototyping Experiment III.1. Research Questions and Hypotheses With the current advent of ubiquitous technologies and devices graphical passwords can be considered as a complementary replacement for traditional passwords in specific situations. The purpose of this experiment is to determine whether one graphical authentication concept is more suitable for ubiquitous environments than the other. As both usability and security need to always be considered when evaluating authentication mechanisms the following hypothesis have been formulated: H1: There is a difference in users’ perception of usability between recognition-based and recall-based graphical authentication mechanisms in ubiquitous environments. H2: There is a difference in users’ perception of security between recognition-based and recall-based graphical authentication mechanisms in ubiquitous environments. As ubiquitous environments can differ significantly based on device and context of use, there are two reasons for using both a paper prototype and an interactive prototype in the evaluation: to gather more issues and to simulate different screen sizes and situation by using a larger paper prototype and a small interactive mobile prototype. III.2. Participants & Equipment Krug believes that prototyping should be done on a Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


889


III.4. Procedure The participants had to use a novel authentication mechanism in order to login to a system in a mobile environment. For each experiment session the participant was randomly assigned either to the paper or to the mobile prototype evaluation. They then proceeded to evaluate both ImagO and ImagePass graphical authentication concepts. The order in which the concepts were presented and evaluated was alternated for every subsequent participant. All of the participants were required to work through two tasks, enrolment and authentication, with both graphical authentication concepts. While prototyping ImagO, on both mediums, the participant was presented with three different images containing multiple elements and was asked to select the preferred image for authentication. If the participant felt a strong dislike for all the images the other three available images were presented as an alternative. Once an image was chosen the participant was then asked to select his authentication key by pointing or tapping on different image sections. There was no limit imposed on the authentication key length. In order to determine the expected area tolerance during password selection and the nature of the clicked areas, after successful enrolment the participant was asked a few follow-up questions to elaborate the selection. Eg. If a person/animal was clicked the participant was asked whether it was expected that tapping different body segments should count towards the same selection or treated differently. In ImagO the authentication task was essentially the same as enrolment as the user had to repeat the authentication key, by interacting with the same image. In ImagePass different interface components were prepared as paper cut-outs for the paper prototype while interface screenshot images were used for the mobile prototype. During enrolment the participant was presented with a selection grid containing 30 images and was asked to choose a password by pointing/tapping a sequence of the available images. If the images were disliked, the participant could click the New Images button which in the paper medium prompted the examiner to change the selection grid image, while in the mobile medium a new screenshot image was loaded onto the screen. After the graphical password selection, the participant was asked for the reasons behind selecting particular images, while the second examiner prepared the authentication grid screenshot image containing selected and decoy images. The grid image was then printed and loaded into the mobile device before proceeding with the authentication task. During authentication the participant was asked to enter the previously chosen password by pointing/tapping the correct images in the authentication grid (Fig. 3).

Fig. 3. Example for the ImagePass mobile prototype

During the think-aloud protocol, the participants didn’t generally confer information about security issues unless they were specifically lead to that direction of thought by the examiner. To initiate the process the examiner would ask a question such as: “Do you feel that the graphical password you have selected is secure?”, “Do you think that someone could compromise your graphical password?” or something similar. In addition, at the end of each prototyping session the examiner asked questions to clarify some observations made during the enrolment or authentication task. III.5. Results Based on the audio recordings each session was transcribed and a content log of participants’ commentaries was created. The comments were aggregated and analyzed for each graphical authentication concept using the prototype medium as an independent variable. There was no significant difference for the number of comments between mobile and paper prototype either for the ImagO or for the ImagePass concept, thus a medium independent data analysis was conducted. Qualitative analysis methods such as defect categories [28] or mobile heuristics [19] have not been specifically defined for authentication mechanisms. Hence, based on content analysis it was decided that comments will be coded in the following four categories: • usability issues, comments that focus on usability deficiencies of the concepts; • security issues, comments that focus on security deficiencies of the concepts; • positive remarks, any affirmative expression for the concepts; • improvement recommendations, comments that express constructive suggestions about possible system improvements. Observing the positive remarks through usability and security perspectives was considered, however many of the comments failed to be classified in any of the categories. A summary of the data is presented in Table I and Fig. 4.



890


600

559

Number of comments

500

400

373

ImagO

300

ImagePass 217 188

200

144

128

93

100

61

56

45

0 Usability Issues

Security Issues

Positive Comments

Suggestions

Total

Fig. 4. Total number of comments grouped according to category and prototype TABLE I NUMBER OF COMMENTS FOR BOTH PROTOTYPE CONCEPTS ACROSS DIFFERENT CATEGORIES Usability ImagO

Security

Positive remarks

Suggestions

Total

N

M

SD

N

M

SD

N

M

SD

N

M

SD

N

M

SD

188

4.95

2.31

217

5.71

2.89

61

1.61

0.91

93

2.45

1.29

543

14.29

4.37

ImagePass

128

3.37

1.74

144

3.79

2.59

45

1.18

0.63

56

1.47

0.89

389

10.24

3.26

Total

316

8.32

3.52

361

9.50

5.12

106

2.79

1.38

149

3.92

2.32

932

24.53

7.10

In general, the 38 participants made 932 comments (M=24.53, SD=7.1). On average the participants made more comments about the ImagO prototype, whereas a one-way ANOVA revealed a significant difference for the number of comments between graphical authentication concepts F(1, 38) = 13.184; p=0.00 < 0.05. These results partially support both hypotheses, showing that ImagO was generally commented more than ImagePass. A total of 316 comments (M=8.32, SD=3.52) regarding usability were made by all the participants. A one-way ANOVA found a significant difference for the number of usability comments between graphical authentication concepts F(1, 38) = 14.350; p=0.00 < 0.05. The participants discussed usability more during the ImagO prototype (M=4.95, SD=2.31), than the ImagePass prototype (M=3.37, SD=1.74). Clustering the comments identified 24 unique usability issues, 14 issues for ImagO (M=0.37, SD=0.67) and 10 issues for ImagePass (M=0.26, SD=0.54). This supports the first hypothesis by showing a difference in the perception of usability in favor of ImagePass. When discussing security, a total of 361 comments (M=9.50, SD=5.12) were made by all the participants. A one-way ANOVA using the number of security comments as the dependent variable found a significant difference between graphical authentication concepts F(1, 38) = 11.297; p=0.00 < 0.05. Security issues were discussed more during the

ImagO prototype (M=5.71, SD=2.89), than the ImagePass prototype (M=3.79, SD=2.59). Clustering the comments revealed 17 unique security issues, 11 issues for ImagO (M=0.29, SD=0.52) and 6 issues for ImagePass (M=0.16, SD=0.77). This supports the second hypothesis by showing a difference in the perception of security in favor of ImagePass. The participants expressed 106 positive comments about the concepts. A one-way ANOVA using the number of positive comments as the dependent variable discovered a significant difference between graphical authentication concepts F(1, 38) = 10.791; p=0.01 < 0.05. More positive comments were mentioned during the ImagePass prototype (M=1.61, SD=0.91), than the ImagO prototype (M=1.18, SD=0.63). As positive comments could not be strictly categorized into usability or security categories these results partially support both hypotheses by showing a positive prevalence for the ImagePass concept. While working on the prototypes the participants made 149 suggestive comments (M=3.92, SD=2.32) on what they considered to be system improvements. A oneway ANOVA using the number of suggestive comments as the dependent variable discovered no significant difference between graphical authentication concepts, although more suggestions were made for the ImagO prototype (M=2.45, SD=1.29), than the ImagePass prototype (M=1.47, SD=0.89).



891


IV.

the basis for the further development of a usable and secure graphical authentication mechanism.

Discussion and Conclusion

Generally, the participants were amused by both concepts for graphical authentication. Regarding medium specific unique usability issues, during the interactive prototype session while entering the graphical password the participants expected some sort of a system response to each selection for both concepts. Participants suggested either screen vibrations or briefly increasing the lighting of the selected area. The lighting proposition was prevalent in the ImagO concept suggested as adding an increased awareness to the user of the clickable area tolerance. This follows to the second medium specific unique usability issue, which was the difference in responses between the paper and the mobile prototype for the tolerance of the selection area. In ImagO, mobile prototype participants expected a larger tolerance and fewer regions when compared to participants working with the paper prototype. This is understandable as essentially the paper medium tentatively presented the user with a larger “screen size”. In addition, during ImagePass enrolment most participants either thought that “the number of available images is too large for the available screen space,” or “the size of the available images is too small”, which in essence refers to the same issue of cluttered screen content. In both prototype concepts during graphical password entry the participants expected additional information as to what they have entered before submitting the password. On the subject of security there were no unique mediumspecific issues. The main security concern of participants was shoulder-surfing, which was expressed strongly on almost every occasion during the paper prototyping sessions. By prototyping both conceptual approaches to authentication it can be safely concluded that recognition-based systems have lower usability and security problems as well as a higher user preference. In addition, the results of this study passively support previous findings [7][13][22], confirming that although a higher fidelity prototyping medium does potentially discover more issues, it doesn't significantly improve the outcome of the process. Nevertheless, it could be suggested that for ubiquitous environments when only one type of prototype can be developed due to project constraints, an interactive prototype would maximize the comprehension of the developed system. As prototypes were compared for only two graphical authentication concepts it could be beneficial to repeat the evaluation for additional concepts in order to further generalize these findings. Summarizing the results it is evident that there is support for both tested hypothesis. There is a difference in users’ perception of both usability and security between recognition-based and recall-based graphical authentication mechanisms in ubiquitous environments. ImagePass, the recognition-based concept, prevails with fewer usability and security issues and was selected as

References [1]

[2]

[3] [4]

[5]

[6]

[7] [8]

[9]

[10]

[11]

[12]

[13]

[14] [15]

[16]

[17]

[18]

[19]

[20]


Snyder, C., 2003. Paper prototyping: the fast and easy way to design and refine user interfaces. Morgan Kaufmann Publishers, San Francisco. Ehn, P., Kyng, M., 1991. Cardboard Computers: Mocking it up or Hands-on the Future. Design at Work, pp. 169–196. Laurence Erlbaum Associates. Preece, J., Sharp, H., Rogers, Y., 2002. Interaction Design. John Wiley & Sons. Kechid, S., Tamine-Lechani, L., Boughanem, M., Drias, H., 2007. Personalizing Information Retrieval in a Distributed Environment. International Review on Computers and Software (IRECOS), 2-2, pp. 98 - 107. Miloucheva, I., Wagner, D., Niephaus, C., Hetzer, D., 2008. UserCentric Identity Enabled QoS Policy Management for Next Generation Internet. International Review on Computers and Software (IRECOS), 3-4, pp. 363 - 374. Nielsen, J., 2003. Paper Prototyping: Getting User Data Before You Code. http://www.useit.com/alertbox/20030414.html (accessed October 2010). Rudd, J., Stern, K., Isensee, S., 1996. Low vs. high-fidelity prototyping debate. ACM Magazine vol. 3-1, pp. 76–85. O'Neill, E., Johnson, P., Johnson, H., 1999. Representations and user-developer interaction in cooperative analysis and design, Human-Computer Interaction 14, pp. 43-91. Virzi, R.A., Sokolov, J.L., Karis, D., 1996. Usability problem identification using both low- and high-fidelity prototypes. Proceedings of the SIGCHI conference on Human Factors in Computing Systems (CHI ’96) pp. 236-243. Sefelin, R., Tscheligi, M., Giller, V., 2003. Paper Prototyping What is it good for?: A Comparison of Paper- and Computerbased Low-fidelity Prototyping. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems – Extended Abstracts (CHI ‘03), pp. 778-779. Walker, M., Takayama, L., Landay, J.A., 2002. High-Fidelity or Low-Fidelity, Paper or Computer? Choosing Attributes When Testing Web Prototypes. In Proceedings of the Human Factors and Ergonomics Society Annual Meeting Proceedings, pp. 661665. Vaidyanathan, J., Robbins, J. E., Redmiles, D. F., 1999. Using HTML to Create Early Prototypes. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems – Extended Abstracts (CHI '99), pp. 232-233. Liu L., Khooshabeh P., 2003. Paper or Interactive? A Study of Prototyping Techniques for Ubiquitous Computing Environments. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems – Extended Abstracts (CHI ‘03), pp. 774-775. Rettig, M., 1994. Prototyping for tiny fingers. In Communications of the ACM vol. 37-4, pp. 21–27. Polson, P., Lewis, C., Rieman, J., Wharton, C., 1992. Cognitive walkthroughs: A method for theory-based evaluation of user interfaces. International Journal of Man Machine Studies vol. 365, pp. 741–773. Nielsen, J., Molich, R., 1990. Heuristic evaluation of user interfaces. In Proceedings of the ACM Conference on Human Factors in Computing Systems, pp 249–256. Dahlback, N., Jonsson, A., Ahrenberg, L., 1993. Wizard of Oz Studies - Why and How. In Proceedings of the International Conference on Intelligent User Interfaces, pp. 193–200. Kjeldskov, J., Stage, J., 2004. New Techniques for Usability Evaluation of Mobile Systems, International Journal of Human Computer Studies, vol. 60-6 pp. 599-620. Bertini, E., Gabrielli, S., Kimani, S., 2006. Appropriating and Assessing Heuristics for Mobile Computing. In Proceedings of the Workshop on Advised Visual Interfaces (AVI'06), pp. 119-126. Hendry, D.G., Mackenzie, S., Kurth, A., Spielberg, F., Larkin, J., 2005. Evaluating Paper Prototypes on the Street. In Proceedings


892


[21]

[22]

[23]

[24] [25]

[26]

[27] [28]

of Conference on Human Factors in Computing Systems Extended Abstracts (CHI '05), pp. 1447-1447. Sá, M., Carriço, L., 2006. Low-Fi Prototyping for Mobile Devices. In Proceedings of SIGCHI Conference on Human Factors in Computing Systems (CHI '06), pp. 694-699. Lumsden, J., MacLean, R., 2008. A Comparison of Pseudo-Paper and Paper Prototyping Methods for Mobile Evaluations. In Proceedings of the International Workshop on Mobile and Networking Technologies for social applications (MONET'2008), pp. 538-457. Biddle, R., Chiasson, S., van Oorschot, P.C. 2011. Graphical Passwords: Learning from the First Twelve Years. ACM Computing Surveys 44-4. Blonder, G.E., 1995. Graphical Password. U.S. Patent 5559961. Lucent Technologies, Inc. New Jersey. De Angeli, A., Coventry, L., Johnson, G., Renaud, K., 2005. Is a picture really worth a thousand words? Exploring the feasibility of graphical authentication systems. International Journal of Human-Computer Studies Volume 63, Issues 1-2, pp. 128-152. Mihajlov M., Jerman-Blazic B., 2011. Memorability, Performance and Perception in Graphical Authentication. Interacting with Computers, Elsevier. Krug, S., 2002. Don’t Make Me Think: A Common Sense Approach to Web Usability, New Rider Publishing. Lindgaard, G., 1994. Usability Testing & System Evaluation: A Guide for Designing Useful Computer Systems. London. Chapman & Hall.

conferences and workshops as a speaker, invited speaker and chair or a member of the programming committees. She has published over 500 scientific papers and discussions of professional studies in domestic and international media, 154 communications on scientific meetings, 15 chapters in scientific books, 6 books from which one was published bythe Computer Association of Great Britain and other 142 scientific contributions. Dragan Tevdovski was born in Skopje, Republic of Macedonia in 1979. He received M.Sc. in Statistics from Faculty of Economics, University of Belgrade, Serbia in 2006, and Ph.D. in Economics from Faculty of Economics, University “Ss. Cyril and Methodius” in 2010. He is coauthor of the two editions of the University text book: “Statistics for business and economics”, part of the team of authors of the scientific monograph “Nobel prize winners in economics 1969-2008” and together with the professors from EU is author of scientific monograph “Labor Market Characteristics in Selected Economies”. He is also author of 23 scientific papers and 3 policy papers. Coauthors of some of the papers are professors from USA and EU. His primary research interests include economic policy, econometrics and labor markets. Dr. Tevdovski is member of National Bank of the Republic of Macedonia Club of Researchers.

Authors’ information 1

Faculty of Economics, Ss. Cyril and Methodius University, Skopje Macedonia. 2 Jozef Stefan Institute, Ljubljana, Slovenia. 3 Faculty of Economics, Ss. Cyril and Methodius University, Skopje Macedonia.

Martin Mihajlov, born in 1979 in the Republic of Macedonia, has received his M. Sc. and Ph.D. at the Jozef Stefan International Postgraduate School in Ljubljana, Slovenia. His area of expertise is human-comuter interaction, specifically, usable security. He is an author of over 20 papers published in international scientific conferences and journals. He has also participated in many TEMPUS, FP6, FP7 and COST Action projects as a researcher or management committee member. Prof. dr. Borka Jerman-Blažič is a full professor at the University of Ljubljana, Department of Economics and is heading the Laboratory for Open Systems and Networks at Jožef Stefan Institute. The Laboratory under her leadership is involved more than twenty years in European Union Framework Program projects in the area of ICT and related field. Borka Jerman-Blažič is an appointed member to UNECE/CEFAT UN (Economic Commission for Europe) group for Internet Enterprise Development, appointed member of eTEN management committee of EU, member of FP7 Programming Committee on Security, Chair of the Internet Society of Europe (www.isoc-ecc.org) in the first mandate (2004-2007), distinguished member of Slovene Society for Informatics, member of the editorial board of the International Journal of Technology Enhanced e-Learning and International Journal on Advances in Internet Technology. She is also Chair of Slovenian Standardisation Committee on ICT as well as chair of the Slovenian chapter of Internet Society and is member of the European ICT Standardisation Board. Borka Jerman-Blažič has been involved in more than 150 international



893


Some Open Problems in Multimedia Digital Fingerprinting Song Tang1, Li Liu2

Abstract – As an important digital copyright protection technology developed in recent years, digital fingerprinting achieves piracy tracing by embedding a unique serial number called fingerprint into each distributed multimedia copy invisibly. When a pirated copy is found somewhere, the embedded fingerprint can uniquely identify the source of the leakage. Although there are many studies on the digital fingerprinting, there are still some problems have not been satisfactorily resolved. There are many open problems regarding the digital fingerprinting technology, both theoretical and practical nature. In this paper, we analyze and discuss some of the major open problems in the field of digital fingerprinting, which are the fingerprint embedding, the collusion attack and the effective distribution of the differently marked copies. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Multimedia Copyright Protection, Digital Fingerprinting, Information Hiding, Secure Content Distribution, Collusion Attack

I.

Introduction

With the rapid development of communication network and multimedia processing technologies, all kinds of multimedia data, such as image, video and audio, are distributed and shared over the Internet widely. This provides the creators and publishers with new opportunities. However, as anything has two sides, technology advances also poses the challenge of insuring the copyrighted multimedia content is appropriately used and distributed. In the last few years, piracy becomes increasingly rampant as the dishonest customer can easily duplicate, modify and redistribute the multimedia content to a large audience illegally. How to protect the related digital copyright, ensuring the proper distribution and usage of multimedia products has become increasingly critical under the open Internet environment. Although cryptographic encryption is a powerful tool for access control and confidentiality protection, the protection usually terminates once the content is delivered and decrypted. Information hiding technology has been developed to overcome the drawback of encryption [1] [2]. As one of the typical applications of information hiding, digital fingerprinting has been utilized to track the customers who use their received multimedia content for unintended purposes, where a unique serial number called digital fingerprint is embedded into each distributed copy invisibly[3][4]. When a pirated copy is found somewhere, the embedded fingerprint can uniquely identify the source of the leakage. Digital fingerprinting is considered as a perfect supplement of encryption and digital watermarking, which has been a hot research topic in recent years. Although there are many studies on the digital fingerprinting, there are still many problems have not been satisfactorily resolved. Manuscript received and revised October 2011, accepted November 2011

894

There are many open problems regarding the digital fingerprinting technology, both theoretical and practical nature. The purpose of this paper is to analyze and discuss the major open problems in digital fingerprinting, which are how to embed the fingerprint information into the multimedia copy robustly, how to defend the collusion attack that launched by some malicious customers and how to deliver the differently marked copies to different customers effectively.

II.

Basic Concept of Digital Fingerprinting

The concept of using digital fingerprinting to protect digital data was first proposed by Boneh and Shaw [5]. Generally speaking, a multimedia digital fingerprinting system consists of two parts, which are distribution sub-system and tracking sub-system. The basic application model of digital fingerprinting system is shown in Fig. 1.

Fig. 1. The basic application model of digital fingerprinting system

In the distribution sub-system, the content provider generates a marked copy by embedding a fingerprint into the original copy. In order to track the dishonest customer (called pirate or traitor) effectively, the fingerprint is Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Song Tang, Li Liu

difference (JND) from human visual system models [10]. Corresponding to each user is a marked version of the content, which can be denoted as:

usually derived from traceability code that can offer protection by providing some form of traceability for pirated data [6]. Once finding an unauthorized copy somewhere, the tracking sub-system is responsible for tracking the customers who use their multimedia product for unintended purposes. In order to achieve the purpose of copyright protection, the used digital fingerprinting scheme is required to have the follow some properties. Robustness. Multimedia data is usually subject to a variety of tampering or signal processing operations in the transmission process, and hence, the used fingerprint embedding algorithm should have a high robustness to the common geometric distortion and signal processing operations. Imperceptibility. This property requires the marked copy has a high quality. In fact, this property is the basic requirement of information hiding technology. Anti-collusion attack. After get the corresponding marked copies, some malicious customers may generate a pirated copy by collusion attack. In the pirated copy, the fingerprint information is destroyed or the signal strength is weakened, which poses serious challenges to pirate tracking. Therefore, the used digital fingerprinting scheme should have good performance of resisting against collusion attack.

yi = x + si

where yi is the marked copy that the customer has got. The spread spectrum watermarking based additive fingerprint embedding algorithm is most common algorithm. In addition, wavelet watermarking technology is also widely used to embed the fingerprints to the original multimedia copy [11]. A major problem in fingerprint embedding is that, most of the embedding schemes take the image as the host signal, and other types multimedia data, such as the audio and video are more complex than the image. Therefore, a embedding algorithm that is suitable for image may not be suitable for video or other data.

IV.

Anti-Collusion Attack Problem

In multimedia digital fingerprinting systems, there is a cost effective attack named collusion attack, where several customers combine several marked copies of the same content to remove or attenuate the original fingerprints to avoid being identified. The customers who contribute to generate a pirate copy are called traitors or colluders. There are mainly two types of collusion attack: linear collusion attack and nonlinear collusion attack. Linear collusion attack is the most common collusion attack, which can be expressed by the following equation:

III. Major Open Problems in Multimedia Digital Fingerprinting In this section, we analyze and discuss three major open problems in the field of multimedia digital fingerprinting, which are the robust fingerprint embedding, the defense of collusion attack and the effective distribution of the differently marked copies.

⎪⎧ Pˆ = λ1 P1 + λ2 P2 + " + λk Pk ⎨ ⎪⎩λ1 + λ2 + " + λk = 1

III.1. Robust Fingerprint Embedding Problem A major issue in fingerprint embedding is that there are many different types multimedia data, a fingerprint embedding algorithm is suitable one of the types but may not be suitable for others. In [7], the authors proposed a fingerprint embedding algorithm that can be depicted briefly as follow. Before embedding, each fingerprint that is generated from some particular traceability codes is modulated into watermarking signal. In this scheme, the used modulation technique is orthogonal modulation [8]. This scheme uses the additive embedding method that is derived from the spread spectrum watermarking proposed in [9], where a watermark signal is added to a host signal. Suppose that the host signal is a vector denoted as x , and the watermark wi that is fingerprint associated with the i-th customer who purchases x . Before the watermarks are added to the host signal, every component of wi is scaled by an appropriate factor, i.e.: si = α wi

(2)

(3)

where Pˆ is the colluded copy and P1 ,P2 " ,Pk are legal marked copies. Average collusion attack is the most common linear collusion attack, where a group of customers collectively obtains an average of their individually fingerprinted copies. Average collusion attack can be expressed by the following equation: ⎧Yˆ = λ1 P1 + λ2 P2 + " + λk Pk ⎪ ⎨ 1 ⎪λ1 = λ2 = " = λk = k ⎩

(4)

In average collusion attack, the marked signals are typically averaged with an equal weight for each customer. Since no colluder is willing to take more of a risk than others, the average collusion attack is widely used. However, it is not the only form of collusion attack. In [12], the authors show some collusion attacks that are not linear. Let Yˆ = l y ,m y ," , m y be the colluded copy, $C$ be the set of

(1)

(

One possibility for α is to use the just noticeable Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

1

2

n

)


895

Song Tang, Li Liu

To reduce the bandwidth usage, the most effective way is to distribute the multimedia content through multicast instead of unicast. Even though the multicast source transmits a single version of the digital content, the multicast receivers can get slightly different versions provided that a marking process is performed by the intermediate routers along the multicast tree or by the receivers themselves. Several different techniques have been proposed to address this problem. According to the different used technical methods, most of these schemes can be divided into four categories: distributed fingerprinting, trusted client devices, unicast-multicast and encryption-based schemes. There are mainly two distributed fingerprinting schemes, which are "Watercasting" [16] and "WHIM" [17]. Trusted devices based schemes can be found in [17] and [18] respectively. The schemes proposed in [19] and [20] are belong to unicast-multicast type. Encryption based fingerprinting is a hot research topic in recent years, and a lot of related schemes have been proposed, such as "Chameleon" scheme [21] and "JFD" (joint fingerprint embedding and decryption) scheme [22]. Although there are many different distribution schemes have been proposed for distributing the marked copies, most of them are impractical for their high communication cost and low security.

colluders' copies and yij be the j-th component of copy Yi , then some typical nonlinear collusion attacks are described as follows: Minimumattack : l yi = min j∈C yij

(5)

Maximumattack : l yi = max j∈C yij

(6)

(

)

Minmaxattack : l yi = min j∈C yij + max j∈C yij / 2 (7)

The most effective method to solve the problem of collusion attack launched by some customers is to use suitable anti-collusion fingerprinting codes to construct fingerprints. After finding a pirated copy, the used anti-collusion code can provide some traceabilities from the pirated data. Design of suitable anti-collusion codes is important in digital fingerprinting field, and a lot of fingerprinting codes have been proposed in recent years. The first systematic binary fingerprinting code was proposed by Boneh and Shaw, in which they combined a base inner code with a random outer code [5]. A balanced incomplete block design based fingerprinting code was proposed in Trappe et al. [7]. This code is called AND anti-collusion code due to the follow property, any subset of $k$ or fewer codewords combined element-wise under AND is distinct from the element-wise AND of any other subset of $k$ or fewer codewords. A projective based fingerprinting code is proposed by Dittmann et al., which is considered to be optimal [13]. In [14], Dittmann et al. proposed an anti-collusion fingerprinting code construction scheme based on finite projective, where the relationship between points and lines is used to represent fingerprints.

V.

Conclusion

As a newly developed digital copyright protection technology, digital fingerprinting has got a lot of attention in recent years. Although there are many studies on the digital fingerprinting, there are still many problems have not been satisfactorily resolved. In this paper, we analyzed and discussed three major open problems in the field of digital fingerprinting. We hope this paper can encourage researchers to further explore more novel and practical schemes on digital fingerprinting for multimedia copyright protection.

IV.1. Effective Distribution Problem The uniqueness of each marked copy poses new challenges to the distribution of the marked copies, especially for video streaming applications where a huge volume of data have to be transmitted to a large number of customers. The simple solution of sending each unique marked copy to the corresponding customer by unicasting is inefficient, as the bandwidth usage grows linearly as the number of customers increases. In addition, the load of the merchant is high, as he must generate and store many different marked copies. Multicast is an efficient transport mechanism for one-to-many communication, which reduces the overall communication cost by duplicating packages only when routing paths to multiple receivers diverge [15]. However, problems arise when attempting to deliver different marked copies using multicast. Because the traditional multicast technology is designed to transmit the same content to multiple users, which cannot be directly applied to fingerprinting applications where different users receive slightly different marked copies. In addition to multicast, peer-to-peer network is also a very efficient distribution method for multimedia data [2].

References [1] [2]

[3]

[4]

[5]

[6]

[7]


M. Barni , F. Bartolini, Data hiding for fighting piracy, IEEE Signal Processing Magazine, vol. 21, n. 2, 28-39, 2004. Defa Hu, Juanjuan Luo, Yan Feng, Copyright protection in P2P networks using digital fingerprinting, International Review on Computers and Software (IRECOS), vol. 6, n. 3, pp. 366-370, 2011. Defa Hu and Qiaoliang Li, Asymmetric fingerprinting based on 1-out-of-n oblivious transfer, IEEE Communications Letters, vol. 14, n. 5, pp. 453-455, 2010. Yan Feng, Xiliang Zeng, Bandwidth efficient anti-collusion digital fingerprinting scheme, International Review on Computers and Software (IRECOS), vol. 6, n. 3, pp. 393-399, 2011. D. Boneh and J. Shaw, Collusion-secure fingerprinting for digital data, IEEE Transactions on Information Theory, vol. 44, n. 5, pp. 1897-1905, 1998.. Defa Hu and Qiaoliang Li, A simple fingerprinting scheme for large user group, Frontiers of Computer Science in China, vol. 5, n. 2, pp. 163-168, 2011. W. Trappe et al., Anti-collusion fingerprinting for multimedia, IEEE Transactions on Signal Processing, vol. 51, n. 4, pp. 1069-1087, 2003.


896

Song Tang, Li Liu

[8]

[9]

[10]

[11]

[12]

[13] [14]

[15]

[16]

[17]

[18]

[19]

[20]

[21]

[22]

Z. J. Wang, M. Wu, H. V. Zhao, W. Trappe and K. J. Ray. Liu, Anti-collusion forensics of multimedia fingerprinting using orthogonal modulation, IEEE Transactions on Image Processing, vol. 14, n. 6, pp. 804-821, 2005. I. Cox, J. Kilian, F. Leighton and T. Shamoon, Secure spread spectrum watermarking for multimedia, IEEE Transactions on Image Processing, vol. 6, n. 12, pp. 1673-1687, 1997. C. Podilchuk and W. Zeng, Image adaptive watermarking using visual models, IEEE J. Select. Areas Commun., vol. 16, n. 4, pp. 525-540, 1998. Hyunho Kang et al., Video fingerprinting system using wavelet and error correcting code, Lecture Notes in Computer Science, vol. 3786/2006, 150-164, 2006. H. V. Zhao, M. Wu, Z. J. Wang, and K. J. R. Liu, Forensic analysis of nonlinear collusion attacks for multimedia fingerprinting, IEEE Transactions on Image Processing, vol. 14, n. 5, pp. 646-661, 2005. G. Tardos, Optimal probabilistic fingerprint codes, Journal of the ACM (JACM), vol. 55, n. 2, pp. 1-24, 2008. J. Dittmann, P. Schmitt, E. Saar, J. Schwenk, and J. Ueberberg, Combining digital watermarks and collusion secure fingerprints for digital images, SPIE J. Electron. Imag., vol. 9, n. 4, pp. 456-467, 2000. I. Brown, C. Perkins and J. Crowcroft, Watercasting: distributed watermarking of multicast media, Proceedings of the First International Workshop on Networked Group Communication, pp. 286-300, 1999. P. Judge and M. Ammar, WHIM: Watermarking multicast video with a hierarchy of intermediaries, Computer Networks, vol. 39, n. 6, pp. 699-712, 2002. B. Briscoe, I. Fairman, Nark: receiver based multicast key management and non-repudiation, British Telecom technical report, June 1999. F. Bao, Multimedia content protection by cryptography and watermarking in tamper-resistant hardware, Proceedings of the 2000 ACM Workshops on Multimedia, pp. 139-142, 2000. T. L. Wu and S. F. Wu, Selective encryption and watermarking of MPEG video, International Conference on Image Science, Systems, and Technology, 1997. H. Vicky Zhao, K. J. Ray Liu, Fingerprint multicast in secure video streaming, IEEE Transactions on image Processing, vol. 15, n. 1, pp. 12-29, 2006. R. J. Anderson, C. Manifavas, Chameleon-a new kind of stream cipher, Proceedings of the 4th International Workshop on Fast Software Encryption, pp. 107-113, 1997. D. Kundur, Video fingerprinting and encryption principles for digital rights management, Proceedings of the IEEE, vol. 92, n. 6, pp. 918-932, 2004.


Department of network media, Hunan Mass Media Vocational Technical College, Changsha 410100, Hunan, China. 2 Center of information, The Third Xiangya Hospital of Central South University, Changsha 410013, Hunan, China.

Song Tang received his M. S. degree from Hunan University in Software Engineering. Now, he is a reseacher at Hunan Mass Media Vocational Technical College.

Li Liu received her M. S. Degree in computer science and technology from the National University of Defense Technology. Now, she is a researcher in center of information, the third Xiangya hospital of Central South University, china.



897


QoS Correction in IMS Networks M. Errais1, B. Raouyane1, M. Bellafkih2, M. Ramdani1

Abstract – The network management, especially monitoring and correction of QoS degradation, is becoming an increasingly difficult task. Indeed the network size evolution and diversity of deployed technologies in new architectures. Particularly in IMS networks, which the management tasks and resource configuration is complex operation. In this paper, a new architecture for management and control IMS networks, based on eTOM specifications. The architecture aims to move toward an independent system for monitoring and correction of QoS degradation automatically in real time. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: IP Multimedia Subsystem (IMS), Enhanced Telecom Operation Management (eTOM), Network Management, QoS Control

I.

is based on eTOM processes translation, modelling and conception as Web Services (WS) exposed by the SOA [8] architecture. The organization of this paper is as follows. Initially we present the service provisioning in both cases IMS networks and eTOM framework. Then, we explain our approach which projects eTOM to IMS for monitoring and correction. Finally, we retrieve performance indicator of approach from test bed for studying it behavior.

Introduction

The telecommunications networks are growing rapidly, with multiple service and entity types. This poses a challenge to operators for an efficient administration and management. As new generation network, the IMS [1] networks offer a new architecture for QoS management and service delivery, which ensure greater accessibility of multimedia services. However, the diversity of access networks supported, and nature of deployed services, requires the introduction of innovative solutions for not only QoS management but service assurance.The 3GPP specifications offer dynamic and flexible scenarios for QoS management in IMS networks. The scenarios support the constraints required by service and resources status. However, the proposed mechanisms are focused just in service provisioning without QoS monitoring or service assurance [2]. Thus, these supply entities are unable to detect QoS degradation or propose a service restoration in real time.In this paper, we propose an autonomous system for monitoring and correcting QoS trouble in IMS network, in real time. This system implements an innovative approach for services monitoring [3] based eTOM [4] processes. With a possibility to enable self-correction and autoconfiguration of IMS networks by introducing mechanisms processes-based to provide adequate solutions in trouble cases. The proposed solutions take into account type of service, resources state and add also business constraints described in SLA [5] for each customer. This allows operators to develop the relationship with customers, and gives an added value for differentiate supply services and assurance. The approach modelling is based SLA scenarios defined in the eTOM [6] framework with flexible projection in the IMS network. Indeed, the new approach

II.

IMS Networks

II.1.

IMS Architecture

The IMS offer a new architecture for network control and service supply. The architecture consists of three layers each of which focuses on a particular aspect: • Service layer: A layer that includes all the services deployed by the operator. Application servers are highlighted in the IMS architecture, which is explained by the emergence of business of the operator which converges more towards the integration of services, to meet customer needs. • Control Layer: This layer contains control entities of IMS access CSCF. And record structures of customer profiles in HSS. The communication between these entities is established based signalling via the protocols SIP [8] and Diameter [9]. • The access layer: It includes the various broadband networks deployed by operator. The IMS architecture supports different types of access networks, wired and wireless, which ensures high availability of services deployed. Such an organization significantly eases new services deployment and integration. The information sharing between access networks and other layers, allows


898


M. Errais, B. Raouyane, M. Bellafkih, M. Ramdani

standardizing client authentication and delivery of service. However, the diversity of access and type of services pose a great challenge for the QoS management. II.2.

level of QoS and security. Similarly, distributed aspect of configuration scenarios, enables a robust resources reservation regardless the access type into networks. However, QoS management entities, as described by 3GPP, focus on service provisioning, without a mechanism for QoS monitoring and troubles correction. For this, we purpose a modeling of an autonomous system for monitoring services in IMS network; while ensuring high reliability of deployed services.

Provision Service in IMS Networks

The 3GPP specifications provide dynamic scenarios for the services provisioning in IMS networks. The scenarios are developed based on two defined entities: • Policy charging and rules function (PCRF): is responsible for identifying the optimal solution, following the QoS constraints required for the type of service requested. • Policy and Charging Enforcement function (PCEF): is responsible for establishing and implementation of identified solution by the PCRF [10]. The PCEF [11] is deployed directly on resources. Service delivery is initiated with a SIP-INVITE request sent by the client to P-CSCF, the request contains the required QoS constraints based SDP [12]. The PCSCF, structures the received request, and includes user identity, before forwarding the request to PCRF, via the Diameter protocol (Fig. 1). The PCRF identifies the optimal solution according to required constraints, and then sends to PCEF a RAA request. The PCEF executes the appropriate script to set the configuration described by PCRF. Once completed the configuration, PCEF notify PCRF, which in turn notifies the P-CSCF.

III. eTOM Framework III.1. eTOM Framework Organization The eTOM Framework present an innovative model to support operators in various fields of telecommunications, including service delivery, assurance and taxation of clients. A model, which is based on several processes divided into sub-processes. Each process describes in precise actions to be performed for each situation.

Fig. 2. eTOM Framework organization

The eTOM framework is organized in such a way as to promote actions on customers, allowing operators to focus on business processes able to retain subscribers. The eTOM is based on business processes notion, which are grouped in the form of horizontal and vertical zones (Fig. 2). The most important area in eTOM is Operation, which include a set of E2E Business processes. The 'Operations' area is traditional heart of the business or service provider (SP) and eTOM Framework. It includes all processes that support client (and network) operations and management. It includes a combination of processes and actions of customer support, including management, provisioning and relationships with partners (Fig. 2). The horizontal and vertical processes groupings constitute a matrix formed by a crossing of several processes from level 2, many derivatives of TOM, which connected to customer and support famous operations (Fulfillment, Assurance, and Billing).

Fig. 1. Service provisioning in IMS Networks

The reception SIP request (BYE, CANCEL) notify service end, indeed the PCSCF must request resource release by a STR message, which includes just session identity. The PCRF sends a RAA request to PCEF, for remove the configuration established and to release resources for other utilization. II.3.

ISSUE III.2.

The exchanged information between QoS management entities is performed to guarantee a high

The SID Model

The SID (Shared Information Data) [13] is a standard



899


data model for structuring involved business entities in operator. It includes several diagrams UML data; each one models an eTOM layer. These diagrams are as follows: • Customer: models different entities associated with the client, including recorded profiles, taxation history and service consumption. • Service: collection of entities describing services available and composition, QoS requirements, protocols and mechanisms necessary for service delivery in best conditions. • Resource: includes entities modelling physical and logical resources deployed by operator.

notification which describes service state is sent through Customer Interface Management process to customer.

Fig. 3. The sequence of correction scenario- Resource step

III.3. Correction Scenario The correction scenario defined in the eTOM aims to identify any deterioration in QoS before correction step, in time optimal. A complete scenario that provides mechanisms needed to identify optimum solution in terms of QoS management. The sequence of the scenario correction runs over several stages, the first is executed in Resource layer with simple solution. If troubles persist, the second step is performed to give a new conception by Service processes implementation.

III.3.2. Service Step The second stage of QoS correction is triggered when the processes of Resource layer is unable to resolve resource trouble. Following, need for information for decision-making available in resource layer. Indeed, the division between the three layers service, Customer and Resource, requires the sharing of information in the SID entities, an organization to ensure a high level of security and a more effective structure reports exchanged. The second step (Fig. 4) is initialized by SPM, after recovery of RTM report. The SPM sends a report containing service status and all actions applied, to Service Configuration and Activation (SCA) process.

III.3.1. Resource Step The first step is to develop a solution based processes belonging to eTOM Resource layer. However, notifications are sent toward monitoring layer in order to prepare more complicated solutions. The correction scenario as defined in eTOM does not describe specific solutions, but it focuses on actions organization to be executed and structuring of exchanged reports, without define neither useful information nor correction mechanism. The correction scenario is initiated (Fig. 3) by Resource Data Collection and Processing (RDCP) process, which is responsible for collecting key performance indicators (KPI) [14] and QoS metrics. The KPIs are structured and transmitted to Resource Performance Management (RPM) that applying appropriate thresholds to identify QoS degradation. If the indicators are considered critical, a report is sent to Resource Trouble Management (RTM), responsible for identifying optimal solution, according to KPIs values and resources state. Notifications are sent to Service layer by RPM and RTM processes. Indeed, simultaneously RPM sends to RTM process, and the both Service Quality Management (SQM) transmits to Service Problem Management SPM for synchronizing their information about resource troubles. The SQM in turn, sends a report on service state to Customer QoS/SLA Management (CQSM) process, which downloads customer profile through process Retention and Loyalty (RL), to identify client importance and his signification in enterprise. If the client is deemed important, so a

Fig. 4. The sequence of correction scenario- Service step

The SCA starts to load services available and QoS constraints required for each service. After, the grouping of all necessary information, SCA identifies the optimal solution depending on established situation. This solution is transmitted to the Resource Provisioning (SP) as new service configuration. The process runs a set of scripts necessary for activation proposed solutions. III.4.

Issue

The eTOM framework includes performance scenarios for supporting operator in various fields of activities, such as the service assurance. To this end, the



900


• Service Problem Management (SPM): its main function is join reports to correct resource level, which includes resources state, actions initiated in first correction phase. Before adding customer list sorted by importance. These reports are used to identify service level solution. • Service Configuration & activation (SCA): It identifies service solution, based on service state, client importance and QoS constraints of current services, in network. Service solution is representing as a new conception with novel resource configuration. • Customer QoS/SLA Management (CSQM): is responsible for identifying client importance, and constraints required in SLA, after loading their profile. • Retention and Loyalty (RL): Its main function is to identify client importance. An importance that takes into account services consumption and taxation history. • Customer Interface Management (CIM): The process includes functionality of communication with clients, when sending notifications in event of QoS degradation detection. • Resource Provisioning (RP): it can perform solutions established by SPM and SCA processes as new resource reservation.

QoS correction scenario, present a great opportunity to evaluate IMS networks performance. However, the eTOM specifications are generic without specification for IMS networks. The correction scenario as defined in the eTOM, does not specify actions for identifying solutions or relevant information to be used for detecting trouble in network. The assembled of eTOM and IMS networks will enjoy advantage of new architecture, while ensuring reliability of deployed services and increase financial revenue. In this work, we propose a new approach for monitoring QoS degradation and correction, in real time. In order to achieve an autonomous system that is capable to correct any QoS degradation, in real time.

IV.

The New Approach Proposed

The projection of eTOM framework in IMS requires a compliance with a precise methodology. A methodology for identifying main steps of modeling and implementation, in particular define a set of interaction entities between eTOM processes and IMS components, plus specification of structures for data backup and choice of technical tools to implementation and deployment. IV.1. eTOM Projection in IMS Network The integration eTOM processes in IMS, requires definition of an action explicitly included in each process. Indeed, general specifications eTOM, requires a detailed explanation of actions appropriate to IMS context. These actions and processes defined the main functions of a system of monitoring. The actions allocated to each process in IMS context are as well: • Resource Data Collection & Processing (RDCP): collect and structure information via resources. The type of information that organizes as report depends on entity. Mainly, RDCP treats and recovers key performance indicators (KPIs) in resources. And services parameters and QoS metrics in real time, via IMS entities responsible service provisioning. • Resource Performance management (RPM): based RDCP report and data, the process allows to apply an initial analysis of KPIs collected, in real time. The analysis is mainly based on thresholds defined for each entity and service. • Resource Trouble Management (RTM): The process will be active in trouble cases. It identifies optimal solution for QoS degradation. A simple solution that takes into account type of service, resource state. • Service Quality Management (SQM): It receives correction reports for preparation of alternative solution-level Service. The process also involved reports preparation, thresholds identification to be applied by the RPM on KPIs.

IV.2. Data Presentation The eTOM processes include actions performed to support a business operator, especially in terms of providing assurance services. The actions and functions require data to model situations and support solutions proposed. Thus, data definition is an important step, which has a large impact on system performance. Indeed, a set of data and information are identified, depending on the type but also their uses. IV.2.1. Network Information Network Information includes any information collected for network, in real time, as performance indicators, current configuration of resources. This information is considered dynamic and changing permanently and more importance, which reflect resource status. On a side view, their needs to support solutions proposed by eTOM processes, and another side for identification of anomalies and degradation of service. This information is classified as well: • IMS Access layer: The information collected in real time via access entities, particularly routers deployed. They include key performance indicators for QoS estimation. And current configuration in terms of QoS management for each resource. • IMS Control layer: The information is intended to control events in IMS networks. They include service



901


type, the customer's identity and tracking service as well as IMs signalization. • IMS Service layer: all information collected via service layer is designed to support monitoring operation, in real time. They include mainly parameters of services, for example in VoD service such as ports, codec, and addresses of streaming server. IV.2.2. Shared Information Data (SID) entities The SID entities are used to model operator composition. Multiple class diagrams are defined following SID model organization, which are as follows: • Resource: It defines the logical and physical resources deployed by the operator: ¾ Physical Resource: Several classes that describe physical network entities, particularly application servers, routers, IMS control entities (CSCF, HSS) and QoS management (PCRF, PCEF). It includes type, configuration of devices and access methods. ¾ Logical Resource: These classes specify the logical resources deployed on physical resources. As signalling protocols (SIP, Diameter), communication and management protocols. • Service: The diagram identifies for each Service, QoS management mechanisms used within routers. It includes two main classes: ¾ Service Facing Resource: It describes how to delivery service, for ensuring required QoS. As type of management QoS (DiffServ [15], IntServ [16]) like traffic conditioning. ¾ Service Facing Customer: This class includes specifications of dedicated service to customer’s importance, and type of service (Platinum, Gold, Silver, and Bronze). IV.3.

Fig. 5. Communication architecture

The diversity of entities involved in service provision has imposed distributed system for monitoring. A choice that aims to reduce execution time and sharing processing loads across multiple levels of abstraction. The interaction between IMS components and eTOM processes, which takes its importance in need for information collection exchanged in service provisioning, is achieved via the tree EJB modules WSResource. The distribution of resource modules, deployment of entities interacting on IMS components is a challenge in terms of synchronization and orchestration. For this, a module WS-Synchronization is implemented; which contains tools of synchronization between resources modules, but also a BPEL-Engine for orchestration between exposed Web Services. IV.4.

System Architecture

The aim of proposed system architecture is implementation of new approach while minimizing deployment costs. In this sense, several levels of abstraction are defined, for load balancing between different components of platform. Levels that are as well: • Resource: contains a set of modules responsible for resource management, and information gathering, such as interconnections entities between IMS, and processes of eTOM Resource layer. • Synchronization: is an intermediate layer between resource layer and others layers, it includes the synchronization entities between modules deployed on resources and others. • Assurance: includes the eTOM processes belonging to Customer and Service layers, in addition to save customer profiles and service QoS requirement. The system architecture (Fig. 6) that contains several EJB modules represents three layers of eTOM Resource, Service and Customer. Thus, three layers are deployed independently, which meets the specifications eTOM. To address the needs for data structuring collected in real time, processes are deployed on dedicated entities that participate in the exchange of information.

Technological Tools

The implementation of new approach requires use of innovative solutions in terms of implementation and deployment. The eTOM specification separate an operator’s in three layers for management: Resource, Service and Customer. Each layer contains a several processes, each one is provided to perform a particular function or operation. To this end, SOA (service oriented Architecture) is presented as the best solution suited to needs explained. Indeed, it allows implementation of eTOM processes as web service, exposed via independent EJB [17] modules; these modules represent overall processes in layer. Communication between web services is established based SOAP [18] protocol, which avoids overloading network when exchanging messages.



902


• WS-Service: It outlines the processes belonging to the eTOM service layer. The module also includes the SID entities in service class diagram. • WS-Customer: The EJB module that includes the processes belonging to the eTOM service layer. IV.5.

The Monitoring Sequence of Platform

The monitoring operations in new approach can be classified in three phases, and are as following: IV.5.1. Phase of Service Detection and Reports Distribution The first phase allows collection of information necessary for monitoring, before sending to resources modules, for activate monitoring mechanisms in real time. This phase is initiated after service delivery detected by IMS module, which identifies client type and transmitted the report drawn up, at synchronization module. This latter, adds parameters of services settings, and sends generated report to Service module. The Service module loads SLA customer, before writing monitoring report identified by session code. This report is sent to resources modules, to initiate a permanent monitoring operation.

Fig. 6. System Architecture

The consolidation of reports written and orchestration between the modules deployed on resources is achieved via a BPEL process, includes in the synchronization layer. Before presenting the EJB modules, it is useful to list the different structures of data backup, which are as well: • Resource-inventory: entity that brings various resources, logical and physical, as well as the current configuration of each resource. • Service-inventory: A structured entity in an XML file that includes available services and the QoS requirements of each type of service. • Customer-inventory: It includes profiles of customers registered, as well as the QoS constraints of each type of client. The EJB modules included in platform are as follows: • WS-Resource: EJB module that contains the eTOM processes belonging to the resource layer. It is deployed on routers to facilitate collection of key performance indicators.This module is responsible for anomalies identification by analyzing KPIs collected. And the application of solutions made by upper eTOM layers. • WS-Resource-IMS: The module that implements the communication functionality with IMS entities, it allows service providing detection in IMS and identity of registered customers. The reports are structured and analyzed by the RDCP process. • WS-Resource-AS: This module is deployed on the application server. It is responsible for the service parameters collection, needed to calculate KPIs. • WS-Synchronisation: The module responsible for synchronization a side between modules deployed on resources, and another side between web services by BPEL [19] process. It also exposes the RTM process responsible for correction level-resource. A choice that aims to minimize the burden on the routers, to avoid overloading the network.

IV.5.2. Monitoring Phase The objective of monitoring phase is service monitoring, to identify any QoS degradation, in real time. This phase is initiated by receiving reports via Resource module. So, RDCP process begins collection of key performance indicators (KPIs). These indicators are analyzed by RPM process, using thresholds defined in report. In case the indicators are considered critical, correction operation is triggered. IV.5.3. Correction Phase The QoS correction operation takes two steps in resource and service. -

Resource correction & Reports distribution

The correction step in resource involves eTOM process belonging to resource layer, to provide a first solution in an optimal time. The solutions that do not take into account the type of service, and focus only on concerned resource. The step is initiated by resource module, after QoS degradation detection, the degradation report transmitted to synchronization module. It notifies the service module, and identifies a solution based on the report and the degradation state of resources, through the RTM process. The service module transmitted the report to the Customer module. It notifies customer of the service state, and actions established for the correction. Simultaneously, the synchronization module, sends the correction report to the resource module, for perform the



903


• Case 2: Concurrent multiple streams are transmitted simultaneously in network using the IPERF [21]tool. In both cases, quality of service and behavior of platform are analyzed and discussed.

reconfiguration script. -

Service correction

The second stage of correction is triggered in failure case of resource solutions, the stage are unable to rectify trouble.After indicators analysis, degradation report is written. A report that includes service status and actions applied in first stage of correction. The synchronization module transmitted report to service module, after inclusion of state of all available resources. The service module identifies flows of various services being provision, before sending Customer module. The latter classify stream according to customer’s importance. Thus, service module identifies optimal solution, taking into account resources state, customer importance and type of service. The solution report is transmitted to resource modules, via synchronization module. So, Resource provisioning process runs scripts necessary for resources reconfiguration.

V.

V.2.

Results

In the first case, service quality is considered good (Fig. 8), reflecting absence of competing flows in network. The collected indicators reflect quality of image, so no monitoring operation is initiated.

Tests and Experimentation

The objective of experimental phase is to validate new proposed architecture, and placing it in actual case of service provisioning in IMS networks. The test bench (Figure 7) includes following entities: • Management server which includes two modules customer and services. • Synchronization server that include synchronization module, BPEL-Engine and PCRF. • Two Edge routers which include resource module and PCEF. • Core router that include IMS entities, deployed by OpenIMS [20], resource module, and PCRF. • VoD application server.

Fig. 8. Image captured in the first case

Fig. 9. Image captured in second case

In the second case, presence of competing flow lead to overloading resources, which explains appearance of pixels on video received by Alice (Fig. 9). Key Performance indicators collected are considered critical, according to SLA requirements. Thus, correction operation is initiated. The outbreak of correction scenario resource-level has improved QoS received. Indeed, resources reconfiguration, as a way to make flow of Alice priority in network, allowing minimize number of lost packets, thus improving significantly quality of image (Fig. 10). However, presence of several competing flows makes solutions established insufficient.

Fig. 7. Test bench

V.1.

Scenarios

The test cases are: • Case 1: At first, customer Alice, who type is medium, are recorded in IMS network, and ask VoD.



904


Indeed, execution time of platform is composite; it depends on number of web service involved in each transaction, but also of number of competing services and the operation inside.

Fig. 10. Image captured in the first case, after correction resource-step

Fig. 12. Execution time of platform

The time collection of key performance indicators is less important, reflecting distributed nature of platform, particularly deployment of resource process. This minimizes time of collection. Similarly, time correction step resource is less than that required that correct step service. A difference that explains by nature of solutions proposed, and number of web services involved in each step. Thus, time correction varies between 1450 ms and 5504 ms for 80 flows. The time is acceptable, given complexity of correction operation. The second criterion evaluated is detection time of service degradation. A criterion that has a significant impact on performance of platform, view, that following detection of degradation correction mechanisms are triggered. For this purpose, two variances are defined to evaluate detection time. Time detection system ( τ s ), and client detection time ( τ u ). These two variances are calculated by the following formulas:

The correction resource-step has improved values of key performance indicators. However, these indicators are considered critical according to SLA of Alice. Thus, correction service-step is initiated. This has improved the quality of service (Fig. 11) received significantly.

Fig. 11. Image captured in the first case, after correction service-step

τ u = τ vd − τ ft

An improvement was due to service solutions, which affects all resources available, providing results more interesting. V.3.

(1)

with: - τ vd : Time of visual detection - τ ft : Time of transmission flows competing

Discussion

τ s = τ sd − τ ft

The new approach to monitoring and correction of IMS networks, has expected these objectives by introducing mechanisms for monitoring and QoS correction, in real time. However, it is important to discuss deploying charges such an approach in terms of execution time and CPU utilization. The execution time is a very important criterion for costs study of the approach. Indeed, an important time for correction back to time spread of deterioration to client. Hence, that will be needed to minimize. Fig. 12 shows schematically execution time of four operations included in correction process.

(2)

with τ sd : Time of passing critical thresholds. Time detection depends on number of current stream in networks but also resource state. Ideally, the detection time of degradation by platform should be lower at user’s level. However, static nature of thresholds used for detection, makes important detection time (900 ms in case of 40 flows).



905


The distributed aspect of proposed architecture significantly increases performance deployment. Especially in terms of execution time and charge resources (CPU usage) (article reference). However, the detection time of degradation remains high, posing a challenge to optimal time at correction operation. To this end, we will focus our future work to integrate a detection mechanisms of degradation and services monitoring [22], based dedicated eTOM processes and new estimation techniques [23], for improved times recorded.

Fig. 13. Time of degradation detection

Hence need to introduce new techniques for identification of damage, in an optimal time. The third criterion was analyzed for study of costs of new approach by CPU consumption in routers. The approach minimizes time of verification by integration modules in routers. However, it is important to ensure that load on routers is acceptable. We use routers are Linux machines (CPU (3.40 GHZ)) and RAM (512 MB).

References [1] [2]

[3]

[4]

[5]

[6]

Fig. 14. CPU consumption according to number of flows [7]

Fig. 14 diagrammatically shows the CPU consumption, based on many competing flows, with and without activation of the monitoring entities. The difference recorded between first and two cases, with and without activation of permanent monitoring entities, is acceptable (25% to 70 concurrent streams). The implementation of corrective mechanisms causes the increased load on resources. However, this charge is acceptable (66% to 70 concurrent streams). These results can be explained by choice to minimize deployment process in resource module, and the use of technology and lightweight optimization algorithms implemented.

VI.

[8] [9]

[10] [11] [12] [13] [14]

[15] [16] [17]

Conclusion

The integration of our approach in the IMS context ensures high reliability of services provided. So, will enable operators to take advantage in supply architecture; which ensuring optimal QoS in different conditions. Indeed, coupling eTOM processes and SOA, is used to migrate to an autonomous business system able to detect at any time QoS degradation and the correct, in optimal time.

[18]

[19] [20]

[21] [22]


Poikselka, M. and Georg M. The IMS: IP Multimedia Concepts and Services, John Wiley & Sons Inc. Chichester, England-2009. BELLAFKIH M., RAOUYANE B., ERRAIS M., RAMDANI M., QoS Management in IMS: Diameter-DiffServ, International Conference on Next Generation Networks & Services, 04-06 June 2010, Rabat, Morocco. Raouyane B., Bellafkih M., Errais, M., Ramdani M., Monitoring IMS Based eTOM-WSOA-BPEL, International Review on Computer and Software (IRECOS), Vol.5. n. 3,pp. 355362, May 2010. ITU-T Recommendation M.3050.3 (2004) SERIES M: Telecommunications Management Network Enhanced Telecom Operations Map (eTOM) – Representative Process Flows (eTOM). Errais, M.; Bellafkih, M.; Raouyane, B.; Ramdani, M, Distributed network Monitoring for IMS Network, The 2nd International Conference on Multimedia Computing and Systems, 7-9 April 2011, Ouarzazate, Morocco. Raouyane, B.; Errais, M.; Bellafkih, M; Ranc, D., SLA Management & Monitoring Based-eTOM and WS-Composite for IMS Networks, The 4th IFIP International Conference on New Technologies, Mobility and Security, 7-10 Feb. 2011, Paris, France. Mark D. Hansen 2007. SOA Using Java Web Services, Prentice Hall PTR. SIP: Session Initiation Protocol, June 2002, RFC 3261 Korhonen, J., Tschofenig, H., Arumaithurai, M. Jones, M., Ed., and A. Lior, "Traffic Classification and Quality of Service (QoS) Attributes for Diameter",RFC 5777, February 2010. 3rd Generation Partnership Project; Evolution of policy control and charging (Release 7), 3GPP TR 23.803 V7.0.0 (2005-09). 3GPP TS 29.210 V6.7.0 “Charging rule provisioning over Gx interface (Release 6)”. 2006-12. SDP: Session Description Protocol, April 1998, RFC 2327 Shared Information/Data (SID) Model System View Concepts and Principles, GB926, Version 1.0, Release 4.0 January 2004. Bellafkih, M., Raouyane B., Errais, M., Ramdani, M. , MOS evaluation for VoD service in an IMS network, 5th International Symposium on Communications and Mobile Network, 2010, vol., no., pp.1-4, 2010, Rabat, Morocco. An Architecture for Differentiated Services, RFC 2475 Integrated Services in the Internet Architecture, RFC 1633 EJB 3.0 Specification: http://openejb.apache.org/3.0/ejb-30specification.html. Latest version of SOAP Version 1.2 specification: http://www.w3.org/TR/soap12, W3C Recommendation (Second Edition) 27 April 2007. Business Process Execution Language Version 2.0.Public Review Draft, 23th August, 2006.http://docs.oasis-open.org/wsbpel/2.0/ OpenIMScore – Open source implementation of IMS Call Session Control Functions and Home Subscriber Service (HSS) http://www.openimscore.org/ IPERF - http://iperf.fr/ Errais M., Raouyane B., Bellafkih M., Ramdani M., Enhanced Telecom Operation Management Scenarios for IMS Networks,


906


International Journal of Next-Generation Networks, Vol.3. n. 2, June -2011. [23] Errais M., Raouyane B., Bellafkih M., Fuzzy Logic for QoS Control in IMS Network, International Journal of Computer Applications, Vol.28. n. 9, Article 7, august -2011.

Authors’ information 1 2

LR@I, FST, Mohammadia, Morocco. Networks laboratory, INPT, Rabat, Morocco. M. Errais received his Master from The Faculty of Sciences, Mohammed V University Agdal, Rabat, Morocco, in 2009. He is currently doing his PhD at FSTM, Mohammedia, Morocco, under the supervision of Prof. Mostafa Bellafkih. His research interests include network management and software Development.

M. Bellafkih had his PhD thesis in computer science from the University of Paris 6, France, in June 1994 and Doctorat Es Science in Computer Science (option networks) from the University of Mohammed V in Rabat, Morocco, in May 2001. His research interests include network management, knowledge management, AI, data mining and database. B. Raouyane is a PhD student at the Faculty of Science and Technology Mohammedia specializing in QoS and Network Management in IMS and NGN. His field of expertise is in computer software and hardware, including network, and performance computing. M. Ramdani had his PhD thesis in Computer Science from the University of Paris 6, France, in February 1994 and Habilitation in Computer Science from the University of Paris 6, France, in June 2001. His research interests include the, knowledge management, AI, data mining and database.



907


A Strawberry Disease Image Retrieval Method Inosculating Color and Textural Features Jian Song

Abstract – A strawberry disease image retrieval method inosculating color and textural features is proposed in order to provide farmers with visualized guidance for disease and insect control and overcome the inaccuracy of the diagnosis expert system for vegetable deseases which relies on words to provide information. A detailed analysis is made on the color features of the strawberry disease images, which are preprocessed in HSV color space which is in accord with visual features of human eyes. 4×4 image feature matrix is constructed when the image mean value, variance, measure of skewness, kurtosis, and energy are extracted as the color feature value while the image contrast, texture consistency and the relevance of entropy and grayscale as the textural feature value. After they are normalized by Gaussian normalization method, Mahalanobis Distance is adopted for similarity measurement. A strawberry disease image retrieval method inosculating color and textural features is developed with C++ programming under Visual C++6.0 development environment. It is indicated by test results that this retrieval method has a preferable recognition effect with a Precision Ratio of 65% and a recall ratio of 83%. When this algorithm is applied to the diagnosis expert system for strawberry diseases, its robustness will be greatly enhanced to meet the requirements of disease diagnosis. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Strawberry, Disease Image, Content based Image Retrieval, Color Features, Textural Features

I.

However, most of the current expert systems obtain such information as growing environment and plant disease signs offered by the vegetable growers in text format, which varies from information providers, thus the diagnosis accuracy is affected [7]. When the content-based image retrieval function is added to the expert system, so long as the vegetable growers provide pictures online, the expert system can retrieve in the standard disease image data bank and display successively the matching standard pictures in accordance with the matching degree of the standard disease image and the picture provided by the growers who can acquire the disease information by checking those most similar standard disease images, thereby diagnosis accuracy is improved. The content-based image retrieval method (CBIR) is to extract the feature parameters which can characterize the image contents, compute similarity distance between the query image and the target image, with such content features as the image color, texture, form and spatial relationships being the image index, and retrieve according to similarity matching [8]. In recent years, scholars at home and abroad have done a good many researches in the field of image-based retrieval and gained some research achievements [8]. Chen Bingquan, etc., uses graylevel Cooccurrence Matrix (GLCM) description for image retrieval, i.e., first the color gradient is

Introduction

Strawberry, honored as “The Queen of Fruits” and “Ruby in the Fruits”, is rich in gallogen with anticancer and beautifying effects. In recent years, there is a rapid development in strawberry planting in China, from Heilongjiang in the north to Guangdong in the south, from Shandong in the east to Xinjiang in the west, with a cultivated area of 100,000 hectares, a total output of 1,880,000 tonnes, and an export amount of 100,000 tonnes, which enables China to be the largest strawberry producing and exporting country in the world [1] [2]. However, the spread of the concomitant diseases is one the rise, with a morbidity of over 30% per year nationwide, which causes tremendous losses to the strawberry growers [3] [4]. Moreover, it seems to be a tendency that there are more difficult miscellaneous diseases that are hard to treat, which becomes one of the important factors to restrict the development of strawberry planting in China[5].In order to solve the contradiction between the frequent diseases and lack of domain expert for the timely diagnosis and prevention and cure, many scientific research institutions at home and abroad have developed vegetable disease diagnosis expert systems, which can be used to imitate human experts for remote diagnosis of vegetable diseases and prevention solution by vegetable growers [6].


908


Jian Song

computed in RGB color space, then GLCM is used to describe the image features [10]. Chen Guilan, etc., proposes a new searching algorithm with a comprehensive use of color, texture, and shape features which greatly improves retrieval accuracy and efficiency through setting reasonable feature database [11]. Zhang Huawei, etc., puts forward a new method combining color with shape, which adopts improved Canny operator to perform image segmentation and then extract 7 invariant moment feature components, and which adopts CBIR model of image retrieval to combine the color feature components with the shape feature components, so as to implement an image retrieval system [12]. The domestic and overseas literatures synthesized, CBIR technology has achieved remarkable results, but there still many exist problems demanding prompt solution. There are distinctive characteristics in the course of retrieval when using different types of features to describe different content properties. The method which realizes image retrieval by single visual feature has poor commonality and bad retrieval result. In addition, the stand or fall of image segmentation has an immediate impact on the quality of the extraction of the shape features. In view of the above questions, a new retrieval method inosculating color with texture is proposed to implement strawberry disease image retrieval.

II.

S = 1−

V=

II.2.

(3)

Extraction of Color Feature

p = p ( H ,S ,V )

(4)

Since the digital image is discrete data, the extracted image feature is also discrete data. Therefore, it is easy to calculate their probability distribution. The color features of the strawberry disease image extracted in this paper include: mean value µ , skewness S , kurtosis K , and energy E . Given that the size of the digital image is M × N pixels, the computing formulas for the features are:

In content-based image retrieval system, color is an important descriptor which can simplify object extraction and classification. The expression of the color feature depends on the color model used. So it is vital to choose the appropriate color expression pattern, i.e., the color model. HSV model is a frequently used color perception model, which uses three attributes of color to describe color, according with people’s custom. The transformational relation between HSV model and RGB model is as follows:

( R − G )2 + ⎡⎣ R − B ( G − B )⎤⎦

( R + G + B)

In HSV color space, the color of one point in the image can be represented with a three-dimensional quantity (H, S, V), whose characteristic is usually stated with probability of fish disease image color obtained through statistical method. The probability can be expressed as:

Selection of Color Model

H = arccos

3

Fig. 1. Average response time per number of sites

Color feature is an important visual property of the strawberry disease image. Compared with other properties, color feature has a relatively definite definition, is insensitive to changes of rotation, translation or scale, and shows strong robustness. Therefore, it is widely valued and studied in image retrieval. Color-based retrieval becomes the most fundamental method for the existing content-based image retrieval systems.

1 ⎡( R − G ) + ( R − B ) ⎤⎦ 2⎣

1

(2)

where: H stands for hue, S for saturation, and V for value. It can be seen from the above formula that the transition from RGB to HIS is simple and clear as the three attributes of color is used to describe color in this color model. This research adopts HSV space considering that HSV space can better reflect human’s visual characteristics of observing color and the calculation amount is not so large. HSV color space grayscale is shown in Fig. 1.

Selection and Extraction of Color Feature

II.1.

3 nim ( R,G,B ) R+G+ B

M

µ=

N

∑∑ x ( i, j ) i =1 j =1

M ×N S=

K=

(1)


µ3 µ

3

(5)

(6)

2

µ3 µ22

(7)


909

Jian Song

M

N

E = ∑∑ { x ( i, j )}

2

grayleve. With regard to the fine-textured region, its value of the graylevel cooccurrence matrix is dispersed about, so the corresponding con value is larger. The computing formulas for the con is:

(8)

i =1 j =1

III. Selection and Extraction of Textural Features

CON = ∑∑ ( h − k ) mhk 2

h

Textural features describe surface properties of the image or its corresponding scenery. Unlike the color features, textural feature is not features based on pixels, which needs statistical calculation in the area containing several pixels. Although it is unable to obtain high level image content merely by using textural features, for texture is only an external feature of the object and cannot completely reflect essential attributes of the object, textural features can be used to make quantitative description to a certain degree of the spatial information in the image, which can be regarded as approximate repetitive distribution of some approximate shapes. The difficulty of textural description lies in that it is closely related to the object shape and daedal object shapes and nesting distribution make it extremely hard for textural classification. In image retrieval of strawberry disease, this regional property possesses significant advantages, as it is unlikely to unable matching because of regional deviation. In this paper, the method of graylevel cooccurrence matrix (GLCM) is mainly utilized to achieve image retrieval.

(9)

k

(2) Texture uniformity ( ASM ) This is a measurement of gray level distribution uniformity. When the element value distribution of the graylevel justice matrix is concentrated nearby the leading diagonal, the corresponding ASM value is larger, i.e., graylevel is uniformed, which is regional uniformity. Conversely, ASM value is smaller. The computing formulas for the ASM is: ASM =

∑∑ mhk 2 h

(10)

k

(3) Entropy ( ENT ) Entropy indicates the texture non-uniformity or complexity degree in the image. When the numerical values of various mhk in the graylevel cooccurrence matrix are not significant and comparatively dispersed, ENT value is larger; Conversely, when the numerical values of mhk are comparatively concentrated, ENT value is smaller. The computing formulas for the ENT is:

III.1. Definition of Graylevel Cooccurrence Matrix One of the important textural features is the spatial correlation between the graylevel spatial distribution characteristics in local region and the pixel positions, because the simultaneous joint frequency distribution of the two graylevel pixels at a distance of （ ∆x, ∆y ） in the image can be expressed with graylevel cooccurrence matrix. If the image graylevel is positioned as level N, the cooccurrence matrix N×N can be expressed as M ( ∆x,∆y ) ( h,k ) , in which the value of element located in

ENT = −

∑∑ mhk log2 mhk h

(11)

k

(4) Correlation between pixel and graylevel ( COR ): ⎪⎧ COR = ⎨ ⎪⎩

(h，k) stands for the occurrence times of the pixel pairs (one with graylevel h and the other with graylevel k) with the distance ( ∆x , ∆y ).

IV.

⎪⎫

∑∑ hkmhk − µ x µ y ⎬⎪ / qx q y h

k

⎭

(12)

Strawberry Disease Image Retrieval Method

III.2. Extraction of Textural Features

IV.1. Fusion of Color and Textural Features

Graylevel cooccurrence matrix is utilized to extract textural features, which is often expressed with the following four feature statistics.

Since color features are pixel-based, it needs statistical calculation for every pixel, which reflects the microscopic feature of the image; while textural features can be used for quantitative description of the spatial information in the image to some extent and be regarded as approximate repetitive distribution of some similar shapes, which reflects the macrofeature. The rational fusion of these two will certainly improve the precision ratio and recall ratio. On account of the above analysis, mean value, variance, kurtosis, skewness of the image component H, S, V , and image contrast, texture uniformity, entropy and graylevel

(1) Contrast ( con ) As for the coarse-textured region, the value of mhk in the graylevel cooccurrence matrix is concentrated nearby the leading diagonal. The value of ( h − k ) is smaller here so that the corresponding con value is smaller, because pixel pairs of coarse texture tends to possess identical



910

Jian Song

correlation are all computed in order to guarantee the realtime of the whole system. Thus altogether 16 strawberry disease image parameters are obtained, which form 4×4 matrix as the image retrieval characteristics. Owing to the differences of the numeric areas and physical significance of each feature component of the feature vector extracted, they are required to be normalized so that each component possesses the same weight when the similar distance is calculated. Gaussian normalization method is one good internal normalization method, with the characteristic that some super-large or super-small element values do not have great influence on the whole normalized element value distribution after normalization. We transform each feature vector to interval [0,1] after normalization. Abbreviation and acronyms should be defined the first time they appear in the text, even after they have already been defined in the abstract. Do not use abbreviations in the title unless they are unavoidable.

The content-based query and retrieval is usually a process of stepwise refinement, where there exists a process of feature adjustment and rematch. When a user searches an object, he uses the input mode provided by the human-computer interface of the demonstration system to form a query condition. Then he chooses the image retrieval mode and makes the necessary pretreatment for the image according to it. The image color space is transformed from RGB space to HSV space. Color-based retrieval requires to The image feature extraction module of the demonstration system acquires the corresponding features of the image to be retrieved and brings the results back to the user. The user judges whether it meets the demand. If not, he may consider changing the image retrieval mode. Start

Retrieval mode select

IV.2. Measurement of Color Feature Similarity Mahalanobis distance is suggested by Indian statistician Mahalanobis to indicate the data covariance distance. It is an effective method for computing the similarity of two unknown samples. Different from Euclidean distance, it takes the relationships of all kinds of features into consideration and is irrelevant to scales. As Mahalanobis distance has better performance in pattern recognition than Euclidean distance, Mahalanobis distance is adopted to measure the color feature similarity of the feature vectors. Mahalanobis distance is used to measure the similarity for the color feature vector. The formula is as follows: di, j = ( xi − x j ) Σ T

−1

( xi − x j )

Image preprocessing Retrieval mode modification Similarity matching

Return retrieval result

satisfaction

End

(13) Fig. 2. Flow chart of the retrieval system

where: - di, j is Mahalanobis distance; -

VI.

xi and x j are respectively vector formed by m

The software environment of the experiment is Windows XP, and hardware is Pentium 4CPU2.8G, 1G RAM, and VRAM 512M. The strawberry disease image database has a total of 100 images, with a resolution of 640×480 pixels. Twelve images are returned for each retrieval to the user in a sequence from left to right and from top to bottom according to their similarity degrees. It needs to judge the retrieval result in the content-based retrieval. Generally, two indexes--Precision ratio and Recall ratio are adopted. Recall ratio primarily means the rate between the amount of the relevant images that the user has retrieved and the amount of all the images related to the target image in the database during one query process. Precision ratio mainly refers to the rate between the amount of the relevant images retrieved and that of the amount of all the images retrieved during one query process. The higher the precision ratio

indicators of No.i and No. J sample; - Σ is the sample covariance matrix. After the distance data is acquired, the first several of the most resembled strawberry disease images are output according to the system requirements.

V.

Analysis of the Experiment Result

Systemic Function and its Implementation

Strawberry disease image retrieval system inosculating color and textural features adopts C++ language for programming under Visual C++6.0 development environment and database for storing information of images and their features. Fig. 2 is the flow chart of the retrieval system.



911

Jian Song

[6]

Su Xianming, Wen Jiwen, Li Daoliang, MDA-based group decision support system for teleconsultation of fish disease,Transactions of the CSAE, vol. 25, n. 2, pp. 240-245, 2008 . [7] Amer N. Abuali, Loai F. Al-Zoua'bi, Bilingual Text-Based Image Retrieval Using PDA's, International Review on Computers and Software (IRECOS), vol. 5, n. 3, pp. 303-308, 2010. [8] A. Sabri, M. Karoud, H. Tairi, A.Aarab, An Efficient Image Retrieval Approach Based on Spatial Correlation of the Extrema Points of the IMFs, International Review on Computers and Software (IRECOS), vol. 3, n. 6, pp. 597-604, 2008. [9] K. Houari, Y. Chahir, M. K. Kholladi, Spectral Clustering and Dimensionality Reduction Applied to Content Based Image Retrieval with Hybrid Descriptors, International Review on Computers and Software (IRECOS), vol. 4, n. 6, pp. 633-639, 2009. [10] B.Q.Chen, H.L.Liu, Image retrieval system based on the combination color texture and shape, Journal of Hunan University of Arts and Science, vol. 21, n. 4, pp. 67-70, 2009. [11] G.L.Chen, S. M. Tian, W. Wang, The Technique ofContent-based Image Retrieval and the Applicationin, Agriculture Journal of Agricultural Mechanization Research, vol. 33, n. 5, pp. 176-179, 2010. [12] H.W. Zhang, J.G. Sun, A Research of Image RetrievalBased on Color and Shape, Computer Simulation, vol. 26, n. 11, pp. 207-210, 2009.

and recall ratio, the better the effect of the searching algorithm. It can be observed from the retrieval experiments of 100 strawberry disease images that color feature-based retrieval method can achieves better retrieval effect with the precision ratio of 65% and the recall ratio of 83% and basically meet the needs for strawberry disease diagnosis. The cause of the errors is mainly that the color feature merely shows the statistical distribution of all kind of colors in the image, but does not contain any color space local information. Furthermore, it requires multiple operations and costs relatively more time when extracting color features, so the algorithm speed needs to be further improved.

VII.

Conclusion

A strawberry disease image retrieval method inosculating color and textural features is proposed in order to overcome inaccuracy of the current strawberry disease diagnosis expert system which depends on text information. Color feature is extracted in HSV color space which is consistent with human eye vision feature so as to improve the retrieval effect. The retrieval method inosculating color and textural features overcomes the disadvantage of measuring overall image distribution with singular feature. It is indicated by experiment results that retrieval method based on color and textural features has a fairly good recognition effect with a precision ratio of 65% and a recall ratio of 83%. It will greatly improve the system robustness and meet the demands of diagnosis when this method is applied to vegetable disease diagnosis expert system.

Authors’ information College of Machinery, Weifang University, Weifang 261061, Shandong, China. Jian Song received the Ph.D. degree in agricultural mechanization engineering from China Agricultural University, China in 2006. Currently, he is a associate professor at Weifang University, China. His research interests include image processing and Agricultural robot technology.

Acknowledgements This work is supported by Shandong Provincial Natural Science Foundation, china (No.Y2008G32) and Shandong Provincial universities Scientific Research Project (No.J09LG53).

References [1]

[2]

[3]

[4]

[5]

T. Z. Zhang, T. J. Zhou, Segmentation of strawberry image by BP neural network , Journal of China Agricultural University, vol. 9, n. 4, pp. 65-68, 2004. Z.Y. Xie, T.Z. Zhang, J.Y. Zhao, Ripened Strawberry Recognition Based on Hough Transform, Transactions of the Chinese Society for Agricultural Machinery, vol. 37, n. 5, pp. 158-162, 2006. W. C. Zhen, L. Dai, T.L. Hu, Effect of continuous cropping on growth and root diseases of strawberry, Journal of Hebei Agricultural University, vol. 27, n. 5 pp. 68-71, 2004. H.Y. Zhang, L.Wang, S. Jiang, Control of postharvest disease of strawberries by hot-water treatments and its effects on thequality in storage period, Transactions of the CSAE, vol. 23, n. 8, pp. 270-273, 2007. Wang Guirong, Li Daoliang, Lü Zhaoqin, Design and implementation of SMS-platform system for diagnosis of fish diseases, Transactions of the CSAE, vol. 25, n. 7, pp. 130-134, 2009.



912


A Robust Audio Watermarking Scheme Based on SVD and MDCT Ping Su, Haidong Shi

Abstract – A lot of scholars have successfully applied SVD (Singular Value Decomposition) in watermarking algorithm of digital image. But the application research of SVD in audio digital watermarking only appears in recent two years; one major reason is that the audio signal is one-dimensional signal in time domain, which can not directly carry out SVD decomposition for it. At present, classic audio coding compression standards (such as MP3, AAC, AC3, etc) mostly adopt MDCT (Modified Discrete Cosine Transform) to complete the time frequency transform for audio data. A robust audio watermarking algorithm has been proposed in this paper. First, the MDCT coefficient matrix of original audio is built. Then the watermark is embedded by SVD and adaptive strategy is applied during embedding. The experimental results show that the proposed algorithm has good imperceptibility and robustness for low-pass filtering, noise adding and loss compression. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Singular Value Decomposition, Modified Discrete Cosine Transform, Audio Watermarking, Time Frequency Transform

I.

value of a matrix is determinate according to the matrix theory, but different matrixes may have the same singular value. There is also a watermarking method [6], [7] to achieve the purpose of embedding watermark by modifying U and V matrixes, but such research is only at trial sub stage. As SVD method has many good properties, a lot of scholars have successfully applied SVD in watermarking algorithm of digital image. But the application research of SVD in audio digital watermarking only appears in recent two years; one major reason is that the audio signal is one-dimensional signal in time domain, which can not directly carry out SVD decomposition for it. Ozer [8] uses Short Time Fourier Transform (STFT) to transform the original audio signal into matrix form, and then realizes the embedding of watermark with Document [5] Strategy. Zezula [9] and others build the matrix to be decomposed through MCLT(Modulated Complex Lapped Transform) transform, and this algorithm has certain robustness for normal signal treatment operation, while Wangxiao [10] and other people directly section the sampling time domain value of audio and then build matrix, this document does not analyze the effect of section length on SVD decomposition, on the other hand, directly building matrix to be decomposed from sampling time domain value lacks explanation with actual significance. Gao Xingming [11] and other people build matrix by use of the low frequency wavelet coefficient of audio signal, and watermarking signal is embedded into singular value with quantization modulation strategy. Compared with the aforementioned STFT, MCLT and DWT transform, at present, classic audio coding

Introduction

As a branch of information security, digital watermarking technology has become one of the important research topics for multimedia copyright protection and authentication fields, and has been widely used in digital pictures, audio and video works [1]. Audio digital watermarking can be generally divided into two embedding methods, namely, time-space domain and transform domain. Generally speaking, transform domain is superior in the performance of algorithm [2], [3]. Research on watermarking techniques with higher robustness has become a hot spot for the research on digital watermarking, especially the digital watermarking technique based on singular value decomposition (SVD) has excellent performances including increasing the resistance against geometric distortion, thus it has attached much attention in recent years. Gorodetski [4] was the first to apply SVD in information hiding algorithm, whose primary core is to embed the secret information through slightly modifying image matrix and SVD decomposing diagonal matrix. Domestic scholars Liu Ruizhen and Tan Tieniu [5] also put forward a new watermarking technology based on SVD long ago. Although this technology also modifies the diagonal matrix after SVD decomposition, it is not to directly modify the singular value of the image matrix; instead, it gets a new singular value matrix to replace the first diagonal matrix by secondary SVD. Then, many literatures use the similar strategy, however, such algorithm has substantive defects, because the singular


913


Ping Su, Haidong Shi

transform for N + K point, and we can get N + K independent transformation coefficients. While carrying out inverse transformation, smoothing treatment is carried out for K overlapping samples to reduce the noise between frames. From the above description, we can find that K overlapping samples will be transformed twice, which shall largely reduce the transforming efficiency of DCT. In order to overcome this shortcoming, Prencen and Bradly [12] put forward revised DCT transform, that is, MDCT, its formula is as follows:

compression standards (such as MP3, AAC, AC3, etc) mostly adopt MDCT to complete the time frequency transform for audio data. Just based on this consideration, an audio watermarking algorithm based on SVD and MDCT is put forward in this paper, which firstly carries out MDCT transform for audio signal, then, carries out SVD decomposition for MDCT coefficient obtained from the transform, and choose some singular values as the objects of watermark embedding.

II.

X (k ) =

Singular Value Decomposition (SVD) and Modified Discrete Cosine Transform (MDCT) II.1.

=

1⎞ π ⎞ N + 1 ⎞⎛ ⎟⎜ k + ⎟ ⎟ 2 ⎠⎝ 2⎠ N ⎠

(2)

where, k = 0 ,1," ,N − 1 , h ( n ) is window is window

Orthogonal transformation can decompose signal into the total weighting of some components with mutual orthogonal, principal component analysis (PCA), K-L transform and SVD all belong to the orthogonal transformation of signal, of which SVD is widely used in application fields such as solving optimization, least square method, and multivariate statistical analysis, etc. Supposing X is the matrix for M × N , matrix rank is r , then, there is orthogonal matrix U for M stage, its column is composed of eigenvectors of XX T , there is orthogonal matrix V for N stage, and its row is composed of eigenvectors of X T X , thus:

function, generally adopting sine window:

⎛ π ⎛ 1 ⎞⎞ h ( n ) = sin ⎜ ⎜ k + ⎟⎟ 2 ⎠⎠ ⎝ 2N ⎝

(3)

The expression of IMDCT is: x (n) =

1 N

N −1

⎛⎛

∑ X ( k ) cos ⎜⎝ ⎜⎝ n +

k =0

N +1⎞⎛ 1⎞π ⎞ ⎟ ⎜ k + ⎟ ⎟ (4) 2 ⎠⎝ 2⎠ N ⎠

n = 0 ,1," , 2 N − 1

(1)

III. The Proposed Audio Digital Watermarking Algorithm

where: ⎛Σ D=⎜ r ⎝ 0

⎛⎛

∑ h ( n ) x ( n ) cos ⎜⎝ ⎜⎝ n +

n =0

SVD

U T XV = D or X = UDV T

2 N −1

0⎞ ⎟ , Σ r = diag (σ 1 ,σ 2 ," ,σ r ) , 0⎠

III.1. Watermark Embedding Algorithm Watermark embedding process mainly includes following procedures: (1) Suppose the original digital audio signal is A = {a ( i ) ,1 ≤ i ≤ L} , L refers to the number of audio

σ i = λi , i = 1, 2 ," ,r eigenvalues λ1 ≥ λ2 ≥ " ≥ λr > 0 are entire nonzero

{

(

)}

eigenvalues of matrix X T X . We call σ i ( i = 1, 2 ," ,r )

collecting points, a ( i ) ∈ 0,1, 2," , 2 p − 1

the singular value of X, and X = UDV T the singular value decomposing formula of X.

amplitude of the i audio collecting point, P represents bit number used for each sampling point. Suppose watermark information is W = {w ( i ) , 0 ≤ i ≤ l} , of

II.2.

is the

which w ( i ) ∈ {−1,1} .

MDCT

(2) Carry out sub-frame treatment for original audio signal, of which, audio of the i frame can be represented as:

Generally, orthogonal transformation is carried out in frames (or in groups), and coding for coefficients after the each frame of orthogonal transformation is always independently carried out too, thus, there is inherent un-continuity at the boundary, which may possibly produce large noise. In order to eliminate and reduce such noise, we usually partially overlap data points of adjacent sub-frames and then carry our corresponding transform, for example, DCT transform uses N samples of present frame and correspondent K 2 samples of the last and next frame to consist N + K samples, then, carry out DCT

⎧ ⎢ L ⎥⎫ A ( i ) = ⎨a ( iK + k ) ,1 ≤ k ≤ K , 0 ≤ i ≤ ⎢ ⎥ ⎬ ⎣ K ⎦⎭ ⎩

(5)

of which, K is the number of sampling points (1152 is ⎡L⎤ chosen for this paper), N = ⎢ ⎥ is total frame ⎣K ⎦ number.



914


(3) Carry out MDCT transform with frame as unit, K F= MDCT coefficients shall be produced for each 2 frame. So for the entire audio, N × F MDCT coefficients shall be obtained, thus constructing a MDCT coefficient matrix X with the size of N × F . (4) Carry out SVD decomposition for X, we can get U, V and D matrixes, and Σ r = {σ 1 ,σ 2 ," ,σ r } is the

IV.

In the simulation experiment, the original digital audio signal adopts WAV audio with single track, 16 bit quantization. Watermark information adopts random sequence with the length of 50 and numerical value of {-1, 1}. In the process of watermark embedding, frame length K is 1152, robust factor η is 0.01. The waveform of audio signal with watermark is shown in the diagram. To verify the validity of the algorithm put forward in this paper, the experimental results of imperceptibility and robustness tests are listed out separately below.

matrix composed of nonzero singular values on the upper left of D matrix. (5) In order to keep the balance of algorithm tangibility and robustness, this paper chooses the first l ( l ≤ r ) maximum singular value (σ 1 ,σ 2 ," ,σ l ) as the object of watermark embedding. (6) The embedding of watermark information w(i) adopts self-adaption method:

σ iw = σ i + η * w ( i ) * σ i , 1 ≤ i ≤ l

IV.1. Imperceptibility Analysis At present, ITU-R BS.1387 standard is mostly adopted to evaluate the imperceptibility of audio watermarking algorithm internationally, and the test tool of this standard is PEAQ [13]. In this paper, PEAQ tool is used to test the distortion degree caused by watermark embedding for original audio, of which NMR and ODG can preferably reflect the distortion caused by watermark information embedding for original audio carrier, the distortion is in inverse ratio to NMR numerical value, and in direct ratio to ODG numerical value.

(6)

of which η is robust regulatory factor, thus, nonzero singular value matrix Σ rw with watermark and diagonal matrix D w are obtained. (7) MDCT coefficient matrix X w with watermark can be reconstructed through U, V and D w : X w = UD wV T

(7) ODG

(8) Through inversing MDCT transform, audio data Aw with watermark can be obtained.

Description

When inspecting audio data A w (might have been attacked) with watermark, it is necessary to list out U, V and Σ r matrixes used in embedding, the specific procedures are as follows: (1) By carrying out sub-frame treatment with the same length and MDCT transform for A w , we can get MDCT coefficient matrix X w with the size of N × F . (2) Diagonal matrix D w and nonzero singular value matrix Σ rw can be obtained through the following formula: 0⎞ ⎟ 0 ⎟⎠

IV.2. Robustness Analysis To assess algorithm robustness, in this paper, the attack based on Stirmark Benchmark for Audio [14] is chosen to carry out robustness test for the suggested algorithm. In Table III, the experimental results attacked by Stirmark Benchmark for Audio v2.0 are listed out (“Failed” represents error rate has surpassed 20%). From Table III, we can see that the watermark algorithm suggested in this paper has good robustness for normal signal treatment such as low-pass, noise adding, and MP3 compression, etc, and can preferably resist the attack methods such as sampling point overturn, normalization and zero passage, etc. at the same time, we have also noticed that the suggested algorithm is very sensitive to the attack

(8)

(3) The extraction method of watermark w ( i ) is shown as follows: w ( i ) =

σ iw − σ i η ∗ σi

TABLE I OBJECTIVE DIFFERENTIATION DEGREE OF AUDITORY PERCEPTION QUALITY 0.0 -1.0 -2.0 -3.0 -4.0 Very Perceptible Slight Imperceptible Noisy and not noisy noisy noisy

In Table II, imperceptibility simulation test data of this algorithm is listed out, from the data in the table, we can see that when robust control factor is below 0.1, NMR is below -15 (it is generally thought that when NMR is above 1.5, it represents audio is clearly disturbed), ODG can be guaranteed to be within 0 - -1, thus, the watermark embedded by this algorithm can be considered imperceptible.

III.2. Watermark Extraction Algorithm

⎛ Σ w D w = U −1 X wV = ⎜⎜ r ⎝ 0

Simulation Analysis

(9)



915


methods such as FFT, echo and smoothing, etc, the error rate of watermark is high because such attack can cause the change of significant features or components of audio,

η

TABLE II PEAQ AUDIO QUALITY EVALUATION TEST DATA 0.02 0.05 437.784 437.068 437.784 437.063 -29.4155 -21.7597 2.07552 4.65543 -0.0514135 0.106176 0.136078 0.163523 1.98694 4.43306 3.84451 8.75351 0.0385359 0.0876391 0.245201 0.866725 0 0.017094 0.050 -0.340

0.01 437.599 437.599 -34.0275 1.03617 0 0.133355 0.987002 1.92142 0.0179529 0.0140662 0 0.109

BandwidthRefB BandwidthTestB Total NMRB WinModDiff1B ADBB EHSB AvgModDiff1B AvgModDiff2B RmsNoiseLoudB MFPDB RelDistFramesB ODG

the watermark embedding objective in this paper is just main component of audio signal, which has high robustness.

TABLE III ROBUSTNESS TEST DATA OF ALGORITHM UNDER THE ATTACK OF STIRMARK FOR AUDIO V0.2 Attack type BER(%) Parameter AddBrumm_100 0 AddBrummFreq=55 AddBrumm_1100 2 AddBrummfrom=100 AddBrumm_2100 4 AddBrummto = 4100 AddBrumm_3100 6 AddBrummstep=1000 AddBrumm_4100 10 AddNoise_100 4 AddNoise_300 5 Noisefrom=100 AddNoise_500 8 Noiseto=1000 Noisestep=200 AddNoise_700 10 AddNoise_900

16

Compressor

12

ThresholdDB=-6.123 CompressValue=2.1

AddSinus

4

AddSinusFreq=900 AddSinusAmp=1300

Flippsample

0

Period=100 FlippCount = 1 FlippDist=60

Normalize Nothing RC_LowPass Stat1 Stat2 ZeroCross Amplify Echo

0 0 2 2 0 0 Failed Failed

V.

0.08 437.411 437.411 -17.5925 7.52487 0.551572 0.19366 7.30521 14.8586 0.167613 0.996812 0.0626781 -0.883

0.1 437.669 437.669 -16.2878 8.98001 0.706857 0.206473 8.64727 19.4477 0.228473 0.999958 0.0740741 -1.199

Acknowledgements This work was supported by The National Natural Science Foundation of China. (60671037).

References [1]

Cox, I. J., Kilian, J., Leighton, F.T., Shamoon, T., NEC Res. Inst., Princeton, N.J., Secure spread spectrum watermarking for multimedia, IEEE Transactions on Image Processing, vol. 6, n. 12, pp. 1673-1687,1997. [2] Lixia Zheng, JunXiao, YingWang, Zhiqiang Yao, Wavelet self-adaption digital audio watermarking algorithm resisting amplitude compression, Computer engineering and application, vol. 46, n.15, pp. 96-98, 2010. [3] Yueqiang Li, Transparent and robust algorithm of DCT domain digital audio watermark, Computer engineering and application, vol. 46, n.3, pp. 84-86,2010. [4] Gorodetski, V., Popyack, L., Skormin, V., Samoilov, V., SVD-based approach to transparent embedding data into digital Images, Proceedings of the International Workshop on Mathematical Models and Architectures for Computer Network Security, Lecture Notes in Computer Science (Page: 263-274 Year of Publication: 2001 ISBN: 3-540-45116-1_26). [5] Liu, R. Z., Tan, T. N., SVD-based watermarking scheme for protecting rightful ownership, IEEE Transactions on Multimedia, vol. 4, n.1, pp. 121-128,2002. [6] Di, Y., Liu, H., Raminen, A., Sen, A., Detecting hidden information in images: A comparative study, Proceedings of Workshop on Privacy Perserving Data Mining (Page: 24-30 Year of Publication: 2003 ISBN: 3-540-45116-1_26). [7] Chang, C. C., Tsai, P., Lin, C. C., SVD-based digital image watermarking scheme, Pattern Recognition Letters, vol. 26, n.10, pp. 1577-1586,2005. [8] Ozer, H., Sankur, B., Memon, N., A SVD-based audio watermarking technique, Proceedings of the 7st ACM Workshop on Multimedia and Security (Page: 51-56 Year of Publication: 2005 ISBN: 1-59593-032-9). [9] Zezula, R., Misurec, J., Audio digital watermarking algorithm based on SVD in MCLT domain, Proceedings of the 3st international Conference on System(Page: 140-143 Year of Publication: 2008 ISBN: 978-0-7695-3105-2). [10] XiaoWang, Bosen, BoZhao, Audio blind authentication algorithm based on singular value decomposition and blending, Zhongshan University journal (natural science edition), vol. 43, n. 12, pp. 188-190,2004.

LowPassFreq=9000

ZeroCross=1000 Amplify = 50 Echo = 10

Conclusion

An audio digital watermarking algorithm based on SVD and MDCT transform has been put forward, in-depth study has been proposed for the watermark embedding and extraction process and analysis has been implemented for the imperceptibility and robustness of the algorithm. Experimental results show that this algorithm not only has good imperceptibility, but also can resist most normal signal treatments and attacks.



916


[11] Xingming Gao, Xiangyun Cai, Blind digital audio fragile watermarking algorithm based on singular value decomposition. Audio engineering, vol. 32, n. 4, pp. 66-68, 2008. [12] Princen, J. P., Bradley, A B., Analysis/synthesis filter bank design based on time domain aliasing cancellation, IEEE Transactions on Acoustics, Speech and Signal Processing, vol. 34, n. 5, pp. 1153-1161,1986.

Authors’ information School of Engineering, Zhejiang Business Technology Institute, Ningbo 315012, Zhejing, China. Ping Su is a lecturer at Zhejiang Business Technology Institute. Currenly, her major researh interest is digital watermarkin.

Haidong Shi is a lecturer at Zhejiang Business Technology Institute. Currenly, her major researh interest is digital watermarkin.



917


Recognition of Handwritten Arabic Characters by Using Reduction Techniques Salah M. Al-Saleh1, Salameh A. Mjlae2, Salim A. Alkhawaldeh 3

Abstract – Instance-based learning (IBL) is one of the simplest classification methods that is used for more sophisticated learning techniques such as neural networks. However, IBL has large number of stored instances which leads to the increase in the classification execution time. Therefore, a number of reduction techniques have been developed to overcome this problem. Unfortunately, these reduction techniques were not applied for Handwritten Arabic character recognition. In this paper, we apply four reduction techniques for Handwritten Arabic character recognition. The applied reduction method are Condensed Nearest Neighbour (CNN), Edited Nearest Neighbour (ENN), Instance-Based Learning Algorithm 2 (IB2) and Instance-Based Learning Algorithm 3 (IB3). In addition, a technique that performs a pre-processing of the handwritten Arabic characters patterns is proposed. A resulted data set with a small number of instances related to these patterns is presented. Numerical results presented that the size of the resulted training set is much small than that of the KNN technique with acceptable classification accuracy. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: IBL, Image Pre-Processing, Pattern Recognition, Reduction Methods

I.

The distance metric function or similarity function is used to decide which neighbours are closest to an input instance. The nearest neighbour algorithm uses the Euclidean distance function given as:

Introduction

Pattern recognition is a research area that studies the decision-making operation which classifies the input pattern. It compares its attributes with known attributes in the same class [1], [2]. In the field of machine learning, various recognition techniques have been developed for Handwritten Arabic characters such as neural networks and instanced-based learning (IBL)[1][3]. IBL algorithms simply store some or all of the training instances, and postpone any generalization effort until a new instance is classified. Each instance is assumed to be a set of n attributes. One of these attributes corresponds to the class and the others correspond to the prediction [4]-[6]. K-Nearest Neighbours (KNN) is considered as the simplest type of the IBL methods. It measures the distance metric function to determine the K-nearest neighbours. Then, the class is voted based on the majority instances related to this class. For binary classification tasks, odd values of K are normally used to avoid ties. The most used common values of K are 1, 3 and 5 [7],[8]. For example, as shown in Fig. 1, the training set contains 10 instances, and two classes “ + ” or “ – ”. The shape “?” represents a new instance to be classified. When one nearest neighbour (K = 1) is used, the voted class is “ – ” whereas when five nearest neighbours (K = 5) are used, it is expected that the class “ + ” is voted because four out of five instances related to this class are depicted.

E ( X ,Y ) =

m

∑ ( Xi − Yi )

2

(1)

i =1

where “X” and “Y” are two input instances , “m” is the number of input attributes, and “Xi” and “Yi” are the ith attribute of the input instance and the instance in the training set, respectively [9].

?

Fig. 1. KNN method with the number of neighbours (K) is 1 and 5

Unfortunately, the IBL algorithms suffer from the problem of high storage instances which leads to the increase in the classification execution time. To overcome this problem, reduction techniques have been


918


S. M. Al-Saleh, S. A. Mjlae, S. A. Alkhawaldeh

instance is not the same as the class of the ith instance in T, the ith instance in T is added to S. Then, the added instance is tested with all instances in S to determine which instances are closer to it than the nearest acceptable instance or randomly selected instance. The classification record of these closer instances is updated. As a result, all instances in S with significantly poor classification record are dropped from S. Finally, all nonacceptable instances are removed from “S”. It is worthy of noting that IB3 has higher classification accuracy and much reduction in the number of stored instances than IB2 [10], [11].

proposed to reduce the number of instances in the usage memory with acceptable classification accuracy [10][12]. These reduction techniques were used for several languages. However, they were not used for handwritten Arabic characters. This is due to that Handwritten Arabic characters recognition has attracted relatively less attention compared to other languages. Also, it has no clear classification rules that can be applied because Arabic character can take many sizes and shapes which leads to high complexity [13], [14].In this paper, we propose a new approach that applies four reduction techniques for Handwritten Arabic characters recognition. These reduction method are Condensed Nearest Neighbour (CNN) [15], [16], Edited Nearest Neighbour (ENN) [17], [18], Instance-Based Learning Algorithm 2 (IB2) [5] and Instance-Based Learning Algorithm 3 (IB3) [5].Our approach performs a preprocessing of the Arabic characters patterns and then establishes a data set related to these patterns. Simulation results demonstrated that the number of instances in the training set based on the above reduction techniques is much reduced compared to the KNN technique with acceptable classification accuracy.

II. II.1.

III.

In this section, we present the methodology of our proposed approach. III.1. Handwritten Arabic Characters Pre-processing A character image is transformed into twodimensional binary pattern suitable for computer processing. The binary pattern is then filtered and smoothed from the noise that degrades the image quality. A thinning operation that is also called skeletonization is then applied to convert the pattern into a line drawing. This is required for better character recognition [19]. This pre-processing operation leads to improve the pattern reliability in order to remove the redundant information and to increase the probability to select the desired attributes. Fig. 2 shows a block diagram of the pre-processing stages. At the end of pre-processing stage, the pre-processed pattern can be received as the next stage of the recognition system.

Reduction Techniques

Condensed Nearest Neighbour (CNN)

This algorithm suggests a set of instances “S” as a subset of training set “T”. It begins by randomly selecting one instance from each class of “T” and putting them in “S”. Each instance in “T” is classified using only the instance in “S”. Any instance in set “T” couldn't be classified in set “S” is added to set “S” as new instance [15], [16]. II.2.

Arabic character image

Edited Nearest Neighbour (ENN)

This algorithm starts with letting “S” be the same as “T”, then each instance in “S” is removed if it does not match with the majority of its 3 nearest neighbours [17], [18]. II.3.

Image Representation Binary

Pattern

Smoothing

Instance-Based Learning Algorithm 2 (IB2)

It starts with “S” initially empty, and the first instance in “T” is always added. Then, each instance in “T” will be added to “S” if it is not classified in “S” [5], [10]. II.4.

Methodology and Proposed Approach

Thinning

Instance-Based Learning Algorithm 3 (IB3)

IB3 is a modified version of IB2. In this technique, S starts as an empty set and the first instance in “T” is always added. The ith instance in “T” is compared with all instances in S to obtain the nearest acceptable instance in “S”. If there is no nearest acceptable instance in S, a randomly instance in S is selected. If the class of the nearest acceptable instance or the randomly selected

Alignment

To Recognition Stage Fig. 2. Block diagram of the pre-processing stages



919


III.1.1. Image Representation

The binary pattern matrix is scanned left-to-right, topto-down to detect and examine the noisy pixels. To determine if the pixel is noisy, the following main rules are used: • If a tested bit of 1 occurs along a straight line segment of 0's or vice versa. • If a bit is isolated.

It is used to convert the handwritten Arabic Character image to two-dimensional binary matrix. This step is called digitization and it is necessary for the pattern to be stored and processed in the digital computers. The digitization can be obtained by sampling and quantization. The original signal is sampled at a proper sampling rate, then, quantized by comparing these samples with threshold values. Then, conversion into binary bits “1” or “0” is performed. Binary image has an advantage to reduce the computer storage space. In addition, processing of image can be performed much faster than ordinary images with multi-grey levels. For handwritten Arabic character image, we use a binary matrix with size of 20×20 pixels. To show this, we established a computer software that receives an image and transforms it into binary matrix. Fig. 3 is an example of the output of our software. Note that the handwritten Arabic character pattern is represented only by the bits of “1”.

Fig. 4. 3x3 Sliding Window

As two examples applied by our software, Figs. 5 and 6 show input patterns of two different handwritten Arabic characters with and without smoothing, respectively. As expected, the quality of handwritten Arabic character binary matrix is significantly improved due to the smoothing process. This is noted by correcting most of the noisy pixels (the circled bits) in the two cases.

Fig. 3. Input and computer representation of an Arabic character (‫)ج‬

III.1.2. Smoothing Fig. 5. Input patterns of two Arabic characters

It is a pattern enhancement technique and the first preprocessing operation performed on digitized binary pattern to eliminate the effect of noise which may be presented in the pattern. The noise in the binary matrix is usually produced from the inaccuracy in the imaging device and quantization error in the digitization process [20], [21]. In our approach, we apply a smoothing operation by detecting and marking the noisy pixels on the binary pattern matrix and correcting them through the conversion of these bits to their complements. This smoothing process should be repeated until a very low ignored level of the error is achieved without loss in any important information. The noisy pixels can be detected by using a 3x3 sliding window on the pattern matrix elements [18]. The sliding window represents the tested centred pixel Pij with its 8 neighbours as shown in Fig. 4.

Fig. 6. Smoothed patterns of two Arabic characters



920


III.1.3. Thinning

characters with and without alignment process, respectively. As expected, the output pattern is moved to the top left corner of the binary matrix.

Character thinning is one of the basic pre-processing techniques that can be used to improve the character recognition. Thinning of a binary pattern is generally considered as a process of iterative deletion of pixels along the edge of the pattern until the pattern becomes one line drawing. The process enables much reduction in the pattern data without loss of any important data. Unfortunately, many of thinning techniques are not appropriate for handwritten Arabic characters because they produce a distorted data. To overcome this problem, a modified thinning algorithm has been proposed [21]. In this algorithm, a good connected skeleton of handwritten Arabic character patterns is achieved. Also, it makes the symbol “.” associated with some Arabic characters be represented by a single pixel. The modified thinning algorithm is executed several times over the binary pattern, and in each execution the algorithm marks few pixels to be deleted. The algorithm is terminated when there are no marked pixels. This algorithm scans all left and right edge points, then, the corresponding top and bottom edge points. Finally, sub pattern "." is reduced to a single pixel. In [21], this thinning algorithm was applied for neural networks classification. In this paper, we use this algorithm for reduction method classification. As two examples applied by our software, Figs. 7 and 8 present input patterns of two different handwritten Arabic characters before and after thinning process, respectively. As shown in Fig. 8, the output pattern of handwritten Arabic characters converted into one line, which represents the Arabic characters skeleton.

Fig. 8. Thinned patterns of two Arabic characters

Fig. 9. Input patterns of two Arabic characters

Fig. 10. Alignment of two Handwritten Arabic characters Fig. 7. Input patterns of two Arabic characters

III.2. Handwritten Arabic Characters Recognition III.1.4. Alignment

The IBL methods are the most important classification methods which work to get high classification accuracy. A prominent drawback of these techniques is the memory size due to the large number of stored instances. This leads to increase the classification execution time. Therefore, reduction methods have been proposed to reduce the number of stored instances. In this paper, we

This process is applied as final stage of pattern preprocessing. It aligns the pattern to the top left corner by deleting all rows and columns with bits of 0. This deletion process is repeated until the first row and column contain the bit of 1 at least. Figs. 9 and 10 shows input patterns of two different handwritten Arabic Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


921


use four reduction techniques for handwritten Arabic characters. This paper aims to examine the ability of these reduction methods to recognize handwritten Arabic characters. As expected, much less memory size is achieved with acceptable classification accuracy. To do this, we built a data set containing a large number of isolated handwritten Arabic characters by a large number of people. Smoothing, thinning and alignment are applied to all images. Then, by using conventional IBL (KNN) and four reduction methods (CNN, ENN, IB2 and IB3), the handwritten Arabic characters are classified.

IV.

V.

In our approach, Handwritten Arabic Characters preprocessing was performed. In this process, we apply image representation, smoothing, thinning and alignment for the image. In our approach four reduction methods (CNN, ENN, IB2 and IB3) were used to classify isolated handwritten Arabic characters. Simulation results demonstrated that the reduction methods have much reduced memory size compared to the conventional IBL with comparable accuracy. It was noted that IB3 method has the least memory size compared to other methods. In addition to that all reduction methods decrease the memory size with acceptable classification accuracy.

Simulation Results

In our computer program, we collected 2800 isolated handwritten Arabic character images from random persons as training data set T. This training set contains 100 images for each character. The gathered data were scanned off from the input sheets where the images were stored as two-colour images. The image is then converted into binary matrix of size 20×20. In this matrix, the black pixels represent the written text and take the bit of “1”, while the white pixels represent the background and take the bit of “0”. Smoothing, thinning and alignment are applied to all images. Finally, Classification is performed for the handwritten Arabic characters using KNN (K= 3), CNN, ENN, IB2 and IB3. Fig. 11 compares the reduction methods and the conventional IBL method in terms of memory size and classification accuracy. It is shown that all the reduction methods have much reduced memory size compared to the conventional IBL with comparable accuracy. As expected, the IB3 method has the least memory size compared to all of them whereas ENN has the highest accuracy compared to other reduction methods. It is worthy of noting that although the reduction methods decrease the memory size they have acceptable classification accuracy.

Accuracy Size

IB2

References [1] [2]

[3] [4]

[5] [6] [7]

[8]

[9]

[10]

[11]

[12]

[13] Percentage

100 90 80 70 60 50 40 30 20 10 0 IB3

Conclusion

[14]

[15] [16]

[17]

ENN CNN KNN

KNN and Reduction Methods

[18]

Fig. 11. Memory size and classification accuracy of KNN ( K = 3 ), CNN, ENN, IB2 and IB3 methods


Sergios Theodoridis, Konstantinos Koutroumbas, Pattern Recognition (4th Academic Press, 2008) K. Badie, M. Shimura, Machine Recognition of Arabic Cursive Scripts, Pattern Recognition in Practice, pp. 346-349, North– Holland publishing Co, Amsterdam Holland, 1980. Thomas G. Dietterich, Machine Learning Research: Four Current Directions, AI Magazine , pp. 97-136, Vol. 18, n. 4, 1997. Mohd. Amir, Durga Toshniwal, Instance-Based Classification of Streaming Data Using Emerging Patterns, Information and Communication Technologies, Communications in Computer and Information Science, Vol. 101, part 1, pp. 228-236, 2010. D. W. Aha, D. Kibler, M. K. Albert, Instance-Based Learning Algorithms, Machine Learning, Vol. 6, pp. 37-66, 1991. K. El-Hindi, Early-Halting Criteria for Instance-Based Learning, Proceedings of the AICCSA, Tunisia, 2003. T. M. Cover, P. E. Hart, Nearest Neighbor Pattern Classification, IEEE Transactions on Information Theory, Vol. 13, pp.21-27, 1967. Yong Zeng , Yupu Yan, Liang Zhao, Pseudo nearest neighbor rule for pattern classification, Expert Systems with Applications, Vol. 36, issue 2, part 2, pp. 3587-3595, March 2009. D. R. Wilson, T. R. Martinez, Improved heterogeneous distance functions, Journal of Artificial intelligence Research (JAIR) , Vol.6, n. 1, pp. 1-34, 1997. D. R. Wilson, T. R. Martinez, Reduction Techniques for Instance Based Learning Algorithms, Machine Learning Journal, Vol. 38, n. 3, pp. 257-286, 2000. D. W. Aha, Tolerating Noisy, Irrelevant and Novel Attributes in Instance-Based Learning Algorithms, International Journal of Man- Machine Studies, Vol. 36, pp. 267-287, 1992 . Henry Brighton, Chris Mellish, Advances in Instance Selection for Instance-Based Learning Algorithms, Data Mining And Knowledge Discovery, Vol. 6, n. 2, pp. 153-172, 2002. H. Al–Yousefi, S. Upda, Recognition of Arabic Characters, IEEE Trans, Patt. Anal. Mach. Intll. Vol. 14, n. 8, pp. 853-857, 1992. A. Amin, H. Al-Sadoun and S. Fischer, Hand-Printed Arabic Character Recognition System Using an Artificial Network, Pattern Recognition, Vol. 29, n. 4, pp. 663-675, 1996. P. E. Hart, The Condensed Nearest Neighbor Rule, IEEE Transactions on Information Theory, Vol. 14, pp. 515-516, 1968. Fabrizio Angiulli, Condensed Nearest Neighbor Data Domain Description, IEEE Trans, Patt. Anal. Mach. Intll., Vol. 29, n. 10, pp. 1746-1758, October 2007. I. Tomek, An Experiment with the edited Nearest-Neighbor Rule, IEEE Transactions On Systems, Man and Cybernetics, Vol. 6, n. 6, pp. 448-452, June 1976. R. Alejo, J. M. Sotoca, R. M. Valdovinos, P. Toribio, Edited Nearest Neighbor Rule for Improving Neural Networks Classifications, Advances In Neural Networks , Lecture Notes in Computer Science, Vol. 6063, pp. 303-310, 2010.


922


[19] I. Abuhaiba, S. Mahmoud, R. Green, Recognition of Handwritten Cursive Arabic Characters, IEEE Trans. Patt. Anal. Mach. Intell. Vol. 16, n. 6, pp. 664-672, 1994. [20] S. Unger, Pattern Detection and Recognition, Proceedings of the IRE, Vol. 47, pp. 1737-1752, Oct., 1959. [21] R. Al-Waily, Study on Preprocessing and Syntactic Recognition of Hand-Written Arabic Characters, M.sc. Thesis, University of Basrah, Sep. 1989.

Salameh A. Mjlae was born in Jordan in 1983. He received the B.S degree in information technology from Al - albayt University, Mafraq, Jordan in 2006, the M.S degree in information technology from International University of the Arab Academy for Banking & Financial Sciences, Amman, Jordan in 2008. Since Sept. 2009, he has been as a lecturer with the Department of Computer Science, Zarka University College, Albalqa Applied University, Jordan. His current research interest lies in the areas of pattern recognition and information security.


Department of Computer Science, Zarqa University Collage, Albalqa Applied University, Jordan. 2 Department of Computer Science, Zarqa University Collage, Albalqa Applied University, Jordan. 3 Electrical Engineering Department, Faculty of Engineering Technology, Albalqa Applied University, Jordan.

Salim A. Alkhawaldeh was born in Jordan in 1971. He received the B.S degree in electrical engineering from the University of Jordan, Amman, Jordan in 1994, the M.S degree in electrical engineering /communications and electronics from Jordan University of Science and Technology, Irbed, Jordan in 1999, and the Ph.D. degree in electrical engineering/wireless communications from Concordia University, Montreal, Canada in 2005. Since Sept. 2005, he has been with the Electrical Engineering Department, Albalqa Applied University, Jordan, as assistant professor. His current research interest lies in the areas of digital wireless communications, MIMO transmission systems, space-time coding, OFDM technology and pattern recognition.

Salah M. Al-Saleh was born in Saudi Arabia in 1977. He received the B.S degree in computer science from Applied Science University (A.S.U), Amman, Jordan in 1999, the M.S degree in computer science from Al - albayt University, Mafraq, Jordan in 2003. Since April 2002, he has been as a lecturer with the Department of Computer Science, Zarka University College, Albalqa Applied University, Jordan. His current research interest lies in the area of pattern recognition and artificial intelligence.



923


SVM-kNN Fusion in Vocabulary Tree Method for Specific Object Recognition Amir Azizi1, Sattar Mirzakuchaki2 Abstract – Object recognition is a crucial task in computer vision. In recent years, popularity of object recognition techniques which are based on local features has increased among computer vision researchers. This is due to the robustness of these approaches to real scenes challenges such as viewpoint changes, occlusion, and illumination variations. Vocabulary tree is one of the object recognition techniques based on local features and is used in robotics due to its speed. Thus increasing the accuracy of this method becomes important. In this paper, in addition to evaluation of vocabulary tree’s accuracy for specific object recognition on a difficult dataset we show that the fusion of SVM and kNN classifiers in vocabulary tree leads to an increase in accuracy. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Local Features, Specific Object Recognition, SVM-kNN Fusion, Vocabulary Tree

kNN

k-Nearest Neighbor

Applied weight to vote for -th nearest neighbor Test image’s histogram

SVM

Support Vector Machine

Index of last object in training set

SIFT

Scale Invariant Feature Transform

SVM classifier

DoG

Difference of Gaussian

kNN classifier

RBF

Radial Basis Function kernel

Fusion of two classifiers SVM and kNN

MSER

Maximally Stable External Region

Output label of SVM classifier for

Branch factor of tree Number of nearest neighbors in kNN

Output label of nearest object to the test image in kNN classifier Output label of the second nearest object to the test image in kNN classifier True positive, number of images of one object in test set recognized correctly False negative, number of images of one object in test set recognized incorrectly

Nomenclature

,

, ,

Number of levels in tree Number of descriptors for -th training image in node Number of images in training set Number of training images which have at least one descriptor in node Weight of node

I.

Value of -th histogram bin of -th training image without normalization Value of -th histogram bin of -th training image after normalization Number of descriptors for test image in node i

Introduction

Change of objects scale, viewpoint changes, illumination variations, background clutter and occlusion are some of the difficulties in object recognition task that lead to the utilization of local features in computer vision. A local feature is an image pattern which differs from its immediate neighborhood. Local features can be points, edge segments or small image regions. T. Tuytelaars and K. Mikolajczyk described and compared local feature detectors in [1]. DoG (Difference of Gaussian) [2], Harris Affine and Hessian Affine [3], [4], MSER (Maximally Stable External Region) [5] are some of the most popular detectors. After detecting local features, some measurements are taken from a region centered on each local feature and converted into

Value of -th histogram bin of test image without normalization Value of -th histogram bin of test image after normalization Number of histogram bins Set of distances of the k nearest neighbors Distances of the -th nearest neighbor Set of weights applied to each vote


924


A. Azizi, S. Mirzakuchaki

are introduced in section 3. In section 4 we present our experimental results. A conclusion is presented in section 5.

descriptors or vectors. Detectors which extract points as local features are named interest point detectors and their descriptors are named interest point descriptors. SIFT (Scale Invariant Feature Transform) is the most popular interest point descriptor which was presented by D. Lowe [2]. The extracted descriptors are used in object recognition methods such as Bag of visual words. In the simplest way in this method, extracted descriptors from training images are quantized via clustering. The created cluster centers are called visual words. Each descriptor of each object or image is compared with visual words and a histogram with bins is made where is the number of visual words. Histograms can be trained with a classifier such as SVM (Support Vector Machine). In the recognition phase, descriptors are extracted from test image or object. Each of these descriptors is compared with extracted visual words resulting in a histogram. This histogram is compared with training histograms using a classifier in the algorithm. Classifier’s output is label of test image. Although accuracy of bag of visual words method is acceptable [6], this method is too slow. To have sufficient computational efficiency, vocabulary tree method is presented by Nister and Stewenius [7]. This method is similar to the bag of visual words but quantization is done by hierarchical clustering which not only is performed very fast in training phase but also each of extracted descriptors in recognition phase is compared with a few number of visual words in the tree instead of being compared with all visual words. Fast operation of vocabulary tree has led to its use in applications with fast performance such as robotic. The classifier in the vocabulary tree method is knearest neighbor (kNN), where k stands for the number of nearest neighbors. In this paper we evaluate this classifier in three cases on CMU grocery dataset in addition to the SVM classifier. We propose a SVM-kNN fusion classifier and show that its accuracy is better than other evaluated classifiers. Although applying SVM to vocabulary tree requires a training phase, we show that the recognition time doesn’t increase much. In this paper we utilize CMU grocery dataset which is a difficult dataset for specific object recognition [8]. Some of CMU dataset’s images can be seen in Fig. 1.

II.

Vocabulary Tree

Object recognition using vocabulary tree contains two steps: construction of vocabulary tree and recognition. Since the classifier in vocabulary tree is kNN, there is no need for a classifier training phase. II.1.

Constructing Vocabulary Tree

For all images in the training set a local feature detector is applied and for each of the local features a descriptor is formed. There are some evaluations for selecting proper detectors and descriptors in vocabulary tree [9]. We have used Dog detector and SIFT descriptor in our work because of their simplicity and acceptable accuracy. DoG interest points are fully scale-invariant and robust to rotation. SIFT descriptor is robust to variations in illumination. SIFT descriptors are 128dimensional vectors. A hierarchical clustering stage is performed after making vectors. In the first level, vectors are grouped into clusters using k-means method with being the named branch factor in the vocabulary tree method. In next level, each of groups of vectors is grouped again into clusters; this hierarchical clustering is performed until the -th level. Thus in the first level there are clusters, in the second level there are clusters. clusters and in the -th level there are

Fig. 1. Some of train and test images of CMU grocery dataset

The paper is organized as follows: section 2 describes vocabulary tree method in details. Evaluated classifiers

Fig. 2. Procedure of constructing vocabulary tree and making training histograms



925


with visual words of one group in each level of tree. While this procedure is progressing a histogram is made is the number of test image’s for the test image. descriptors in node . Weighted value of each bin in the histogram is computed as follows:

Each cluster in vocabulary tree is named node and clusters of -the level is named leaf nodes. Nodes have a cluster center which can be named as visual words. Inverted files are exploited for increasing the speed of recognition and constructing the tree. In the process of constructing a tree for each image a histogram is made whose bins correspond to tree nodes. , is the number of descriptors for IM-TH training image in node . For nodes of tree a weight is considered as follows:

· For normalization:

(1)

·

,

nodes participate in above equations where is the number of histogram bins. Recognition in vocabulary tree is done by the kNN classifier so in this method there is no need for a training phase. In the simplest way i.e. one nearest neighbor, distance of test image histogram with each of training image’s histograms is computed. The training image which has the smallest distance with test image’s histogram is known as the most similar image to the test image and test image takes its label. The distance criterion which is used in vocabulary tree is norm-1 due to its accuracy [7] (Fig. 3). The number of levels in the tree along with the branch factor affects the accuracy of recognition [7]. We use 4 levels and branch factor 9 in our work.

(2)

Normalization of histograms is done as follows: ,

,

∑

,

(5)

∑

Value of histogram bins for each training image is computed as follows: ,

(4)

(3)

nodes participate in above equations where is the number of histograms bins. Thus each training image has a histogram which is used for recognition (Fig. 2).

III. Evaluated Classifiers CMU grocery dataset contains 25 images for each of 10 grocery objects for training. As mentioned before, for each image a histogram is formed so we have 25 histograms for each of 10 objects. We evaluate kNN classifier in three cases. In the first case the centroid of histograms of each object is computed and 1-NN is applied on test image’s histogram. In the second case all histograms are considered and kNN is applied on test image’s histogram (where can take on all values between 1 and 250 except 2). In this case maximum vote of nearest neighbors without any weighting is considered as the label of test image. In the third case all histograms are considered and kNN is applied on test image’s histogram but votes take a weight to give significance to nearest neighbors. The weighting to votes of k nearest neighbor is done in the following way [9]: , Fig. 3. Making test image’s histogram and recognition phase in vocabulary tree

II.2.

1

,…,

(6) , 1

(7)

where is the set of distances of the k nearest neighbors and is the set of weights applied to each vote. Support Vector Machine (SVM) is another classifier which is tested on CMU grocery dataset.

Recognition

In the recognition phase, DoG detector is applied on test image with a descriptor formed for interest points using SIFT method. The formed vectors are compared



926


was described in the previous section. We denote the output of the SVM classifier for test image’s histogram as and output of the kNN classifier as , , where ( , , 1, 2, … , ). is the label of the most similar object is the label of to the test image in SVM classifier, is the label of nearest object to the test image and the second nearest object to the test image in kNN classifier. Output of our classifier is the label which is , , labels most. repeated among

We use Radial Basis Function (RBF) kernel for this purpose and utilize LIBSVM package [10]. III.1. Proposed SVM-kNN Fusion (n is number of histograms bins) be the Let test image’s histogram and 1, 2, . . . , be the label set of be SVM classifier c objects in the training set. Let be kNN classifier, and be fusion with RBF kernel, of two classifiers. We consider kNN classifier in the second case which

TABLE I OBTAINED RECALLS FOR EACH OBJECT Object’s name

kNN in the first case

kNN in the second case with k = 12 (the best k)

kNN in the third case with k = 55 (the best k)

SVM with RBF kernel

can chowder can soymilk can tomatosoup carton oj carton soymilk diet coke hcpotroastsoup juicebox rice tuscan ricepilaf average

6 12 4 10 16 30 48 8 18 78 23

38 30 10 34 10 20 64 4 20 50 28

26 28 4 28 12 26 68 6 16 62 27.6

20 4 24 2 4 8 0 40 18 76 19.6

IV.

tree implementation was done in MATLAB. All tests were done in a P8400 computer with 2.26 GHz and 32 bit operation system.

Experimental Results

CMU grocery dataset has 50 images for each object to test. Object masks are provided for these images. To evaluate five mentioned classifiers in vocabulary tree, we compute recall for each of 10 objects in CMU grocery dataset. Recall is defined as follows: 100

Proposed SVMkNN fusion with k = 12 (the best k) and RBF kernel 40 30 16 34 10 22 64 12 22 72 32.2

TABLE II RUNNING TIME OF RECOGNITION PHASE IN VOCABULARY TREE Utilized Classifier in the kNN in the second case Vocabulary Tree with k = 12 Running Time of Recognition

(8)

is true positive and determines the number of where images of one object in test set recognized correctly. is false negative and determines number of images of one object in test set recognized incorrectly.We form our vocabulary tree with branch factor 9 and 4 levels. Thus we have 6561 leaf nodes. Total number of nodes in our tree is 7380.We evaluate vocabulary tree with classifiers which was described in section 3 i.e. kNN classifier in three cases, SVM classifier with RBF kernel and our proposed SVM-kNN fusion classifier. Table I illustrates recall for each object. The best result for kNN classifier in the second case is reached when the number of nearest neighbors is 12. Also, the best result for kNN classifier in the third case is reached when the number of nearest neighbors is 55. And finally, the best result for SVM-kNN classifier is reached when the number of nearest neighbors is 12. Table II shows running time of recognition phase in vocabulary tree with kNN classifier in the second case and SVM-kNN fusion. As can be seen, applying SVMkNN fusion as classifier increases running time of recognition for just a few milliseconds. Our vocabulary

2.062 (s)

V.

SVM-kNN fusion with k = 12 2.067 (s)

Conclusion

In this work we have proposed a classifier in vocabulary tree method on a difficult dataset for specific object recognition. This SVM-kNN fusion is compared with kNN which is the main classifier in vocabulary tree. kNN is considered in three cases. The results demonstrate that our proposed classifier increases recall of object recognition using vocabulary tree in real scenes.

References [1]

[2]

[3]

[4] [5]


T.Tuytelaars and K. Mikolajczyk, Local invariant feature detector: A survey, In Foundation and Trends in Computer Graphics and Vision,Vol. 3, n. 3, pp. 277-280, 2007. D.G. Lowe, Distinctive image features from scale-invariant keypoints, International Journal of Computer Vision, Vol. 60, n.2, pp.91-110, 2004. K. Mikolajczyk and C. Schmid, Scale and affine invariant interest point detectors, International Journal of Computer Vision, Vol. 60, n.1, pp.63-86, 2004. K. Mikolajczyk and C. Schmid, an affine invariant interest point detector, European Conference on Computer Vision, 2002. J. Matas, O. Chum, M. Urban and T. Pajdla, Robust wide-baseline


927


stereo from maximally stable extremal regions, Image and Vision Computing, Vol. 22, n.10, pp.761-767, 2004. [6] A. Bosch, A. Zisserman, and X. Munoz, Representing shape with a spatial pyramid kernel, In Proceedings of the International Conference on Image and Video Retrieval, 2007. [7] D. Nister and H. Stewenius, Scalable recognition with a vocabulary tree,International Conference on Computer Vision and Pattern Recognition, 2006. [8] E. Hsiao, A. Collet and M. Hebert, Making specific features less discriminative to improve point-based 3D object recognition. International Conference on Computer Vision and Pattern Recognition, 2010. http://www.cs.cmu.edu/~ehsiao/datasets.html [9] A. Ramisa, S. Vasudevan, D. Scaramuzza, R. L. de M_antaras and R. Sieg-wart, A tale of two object recognition methods for mobile robots, International Conference on Computer Vision Systems, 2008. [10] Chih-Chung Chang and Chih-Jen Lin, LIBSVM: a library for support vector machines, 2001. Software available at http://www.csie.ntu.edu.tw/~cjlin/libsvm. [11] R. Hemati and S. Mirzakuchaki, Combining harris interest points and the PHOG descriptor for object categorization, International Review on Computers and Software (IRECOS),Vol. 6, n.5, 2011.

Amir Azizi was born in 1983. He received his B.S. degree in Electronic Engineering from Department of Electrical Engineering, South Tehran branch of Azad University, Tehran, Iran in 2008. He is currently master science student of Department of Electrical Engineering of the Iran University of Science and Technology, Tehran, Iran. His research interests include Robot and Computer Vision, Image Processing, Statistical Pattern Recognition and Intelligent Systems. Mr. Azizi is a student member of IEEE. Sattar Mirzakuchaki was born in 1964. He received the B.S. degree in electrical engineering from the University of Mississippi in 1989 and the M.S. and Ph.D. degrees in electronics from the University of Missouri, Columbia, in 1991 and 1996, respectively. He has been an Associate Professor with the Electrical Engineering Department at the Iran University of Science and Technology, Tehran, Iran since 1996. His current research interests include semiconductor growth and characterization and design of VLSI circuits. Dr. Mirzakuchaki is a member of the Institute of Engineering and Technology (formerly IEE) and a Chartered Engineer.

Authors’ information 1,2

Electrical Engineering School, Iran University of Science and Technology, Tehran, Iran.



928


Energy Efficient Routing Protocols for Wireless Sensor Networks: a Survey K. S. Shivaprakasha, Muralidhar Kulkarni Abstract – Wireless Sensor Networks (WSNs) have become one of the emerging trends of the modern communication systems. They find their applications in various fields like habitat monitoring, home automation, environment monitoring, battle field environment etc. WSNs are different from Mobile Adhoc Networks in the perspective of energy awareness, adaptive communication patterns and the routing algorithms. As the sensor devices are powered by batteries, which cannot be recharged often, the power awareness is one of the major requirements in WSNs. Many energy aware routing protocols have been proposed in the literature. In this survey, an attempt has been made to summarize the various energy aware routing protocols available in the literature and also a comparative analysis of these has been made considering various network parameters like the delay, routing overhead, QoS, type of routing protocol etc. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Wireless Sensor Networks (WSN), Base Station (BS), Cluster Head (CH), Medium Access Control (MAC), Region Head (RH)

I.

Finally as the nodes are equipped with the limited power resource, the routing protocol has to be energy aware. Routing in WSNs has been categorized into various categories. They can be classified as proactive, reactive and hybrid protocols. Proactive protocols are generally called as table driven protocols in which the route to the base station from every node will be determined a priori [4], [5]. These algorithms can be used for less dynamic networks. Whereas in reactive protocols the path is discovered only when it is required thus reducing a lot of routing overhead as compared to proactive protocols. Thus they are also called as ondemand protocols. Finally in Hybrid protocols the features of both proactive and reactive algorithms have been included [6]. Another way of classification is based on the network structure as either a flat or hierarchical routing. In flat routing each of the node acts independently whereas in hierarchical routing, nodes are grouped into clusters with a CH node and all transmissions will be via the CH. Routing protocols can also be categorized as either time driven, event driven or query based [8], [9]. In time driven protocols, nodes will be in active state for a fixed amount of time in a periodical manner and senses the data. In case of event driven, the nodes sense the data only when a considerable change in the entity has occurred. And in case of query based, the base station will send a request for the data when it is required and the nodes will reply to the request [10], [11]. A lot of work has been carried out in the field of routing in WSNs [12], [13], [14]. In this paper a survey has been made on the energy aware routing protocols proposed in the literature. Also a qualitative comparison

Introduction

A Wireless Sensor Network (WSN) is a network of hundreds of small devices called sensors, which are deployed either randomly or uniformly over a geographical area. Each node is capable of sensing a physical entity like temperature, pressure, humidity etc [1], [2]. The sensed information will be conveyed to the base station. There are three ways of communicating the information to the Base Station (BS): by direct communication, via intermediate nodes or using clustering method. The first method is feasible only if the BS is in the close proximity of the sender node. Thus multi hop transmission is used in which the sender has to rely on the intermediate nodes to reach the BS. Alternatively the nodes can also be grouped into clusters with one node being the Cluster Head (CH) and the communication to the base station will always be via CH. Generally sensor nodes consist of a Sensor, Processing unit, transmitter, position finding system and power units. Power units cannot be recharged often and thus the data transmission has to be done with the minimum energy consumption. Routing in wireless sensor networks is different than the IP based routing algorithms, as global addressing schemes cannot be used here [3]. Also in WSNs data gathered is more important than the information about the node that has the data. Thus the routing protocol has to be content based. Also as the number of nodes will be large, it is more likely to happen that more than one node can have the same data. Thus the data aggregation also has to be done in order avoid the redundancy.


929


K. S. Shivaprakasha, Muralidhar Kulkarni

As mentioned in Section I, in case of the time driven model, the node will be active for a certain period of time and senses the data. Whereas in case of Event driven, the node will awake only when there is a significant change in the sensing entity. Finally in query-based model, the BS has to initiate the process by broadcasting request signal to the nodes [22].

of the same has been made over various network parameters like routing overhead, delay, QoS, data aggregation etc. The rest of the paper is organized as follows: Section II deals with the various design issues to be considered for the routing protocol in WSN. Section III summarizes various energy aware routing protocols proposed for WSNs. In section IV a comparative analysis of the protocols has been made. Finally section V gives the concluding remarks of the paper.

II.

II.5.

Node deployment also has an effect on the working of the routing protocol. The deployment of the nodes is dependent on the type of the application. It can be either random or deterministic. In case of random deployment nodes will be scattered over a geographical area in a random fashion, which may lead to the formation of the network in an adhoc manner. Whereas in case of deterministic approach, the nodes are manually placed in some order in which the routing algorithms can be simpler [23].

Routing Challenges in Wireless Sensor Networks

As discussed in Section I, routing in WSN is a challenging task. Wireless medium, limited resource availability, hostile environment pose many restrictions on the routing protocols in WSNs. In this section a brief notes on the design challenges for a WSN is studied [15], [16]. II.1.

Deployment of Nodes

II.6.

Energy Usage

Scalability

As mentioned earlier, the wireless sensor networks consist of hundreds of or even thousands of nodes. Also the nodes can join or leave the network with time. Thus it is desirable to have the routing protocol, which should be capable of accommodating the new nodes without affecting the behavior of the network.

As the available energy in the nodes of a WSN is limited, the proposed routing protocol has to optimally use the available resources. If a node’s battery gets extinct, the node becomes dead which may lead to network partitioning. Thus the energy awareness is directly related to the survivability of the network [17], [18].

III. Energy Aware Routing Protocols II.2.

Data Aggregation

As the energy conservation is a vital issue in the performance of the WSN, many protocols have been proposed considering Energy Awareness. In this section we will be summarizing some of the protocols proposed in the literature.

As the nodes may generate redundant data, the transmission of the same will increase the network traffic, which in turn decreases the throughput. Thus the combining of the data has to be performed which is called as data aggregation [19], [20]. This can be done using various methods like duplicate suppression etc. Data aggregation will results in an efficient routing consuming less energy. II.3.

Low-Energy Adaptive Clustering Hierarchy (LEACH) LEACH is a novel cluster based routing protocol proposed in [24]. In this protocol the CHs are chosen on the rotation basis so that the load distribution amongst nodes is almost uniform. Initially each of the nodes has to decide whether to become the CH in the present round depending on some probability. Cluster members are decided based on the distance from the CHs.

Mobility of the Nodes

Although WSNs are assumed to be stationary for most of the cases, there are some applications where the mobility of the nodes also has to be considered. Thus the routing protocols proposed have to be dynamic to accommodate the changes in the network. Also the event can be either static or dynamic. For example it is dynamic for tracking application whereas it is static in case of monitoring systems [21]. II.4.

Geographical Adaptive Fidelity (GAF) GAF is an energy conservative routing protocol, which is independent of the underlying routing protocol [25]. It conserves energy by identifying and turning off the unnecessary nodes in the network. In GAF the whole network is divided into small grids. The algorithm operates in three phases: Discovery, Active and Sleep.

Data Transmission Model

Geographic and Energy Aware Routing (GEAR)

One of the major issues to be considered in routing in WSNs is when to send the sensed data to the BS. There are various models proposed for the same viz time driven model, event driven model and the query-based model.

GEAR algorithm is an energy aware algorithm, which selects its neighbor based on the energy parameter and geographical information [26]. In GEAR the next hop is



930


CHs are changed periodically based on the residual energy and the degree of the nodes [32]. Nodes are considered to be quasi stationary. Initially the CHs are formed based on their residual energy. Number of clusters is predetermined. Nodes will then join appropriate clusters so as to minimize the transmission cost.

selected as a node N if it is closer to the destination with respect to the learned parameter. Learned parameter includes the energy and the distance parameter. Threshold sensitive Energy Efficient sensor Network protocol (TEEN) Time constraint cannot be relaxed for some critical applications in WSNs. TEEN is a reactive protocol, which manages time sensitive applications [27]. It is a cluster based routing where CH does the data aggregation. By managing the thresholds time criticality can be ensured to time sensitive applications.

Gossip Based Routing (GBR) GBR is an energy efficient protocol, which reduces the flooding traffic [33]. Each node floods the message with some probability.Also a lot of energy can be saved if the nodes enter sleep mode whenever they are not active. In GSP, at the beginning of each period, node will decide whether to enter sleep mode or not with some probability p. Nodes awake at the end of each period and the same process is repeated.

Energy Efficient Routing (EER) Traditional protocols concentrate only on the shortest path but do not take into account the available battery power at each node. In Gradient-Based Routing (GBR), while being flooded, the ‘interest’ message records the number of hops taken. This allows a node to discover the minimum number of hops to the user, called the node’s height. When a node detects that its energy reserve has dropped below a certain threshold, it discourages others from sending data to it by increasing its height [28]. Power-Efficient Gathering Systems (PEGASIS)

in

Sensor

Energy-Aware Routing Protocol (EARP) In the traditional routing algorithms like AODV, the path between the source and the destination will be erased after a certain amount of time, which may lead to frequent route discovery initiations. Whereas in case of EARP the table retains all the paths that are less likely to be expired. Thus the same path can be used, if the communication between the same set of nodes have to be established [34].

Information

PEGASIS is a cluster based energy aware routing algorithm and is a near optimal chain based routing [29]. Each node communicates only with its neighbor but not to the CH directly, which reduces the energy consumption at each node. Chain formation will start from the farthest node to the BS. Chain can be reconstructed whenever a node dies in the chain.

Energy Aware Random Asynchronous Wakeup (RAWE) Nodes which are capable of acting i.e to perform a particular job are called as actor nodes eg: Robot. Actor nodes are provided with more battery backup than the sensor nodes.RAW-E distributes the load among the nodes in the forwarding set in proportional to their remaining energy. RAW –E prefers to use actor node [35].

Energy Band based Routing Protocol for Wireless Sensor Networks (EBRP) EBRP is a stateless routing protocol in which the network is divided into various energy bands and the routing is done based on the energy bands [30]. It focuses on the optimal usage of available energies in all nodes. A virtual tree is formed based on the residual energies of nodes in the network with the nodes of lowest energies forming the leaves. Nodes with the same energy level and in each other’s vicity form the nodes at the same level. In each level, nodes will communicate only with the higher-level nodes. Thus nodes with less energy will not be burdened.

COordination-based Data dissemination for sEnsor networks (CODE) CODE addresses the situations where the BS is mobile and the sensors are stationary [36]. It is assumed that all nodes know their geographic information. CODE involves three phases: data announcement, query transfer and data dissemination. SInk cluster – based data Dissemination for sEnsor networks (SIDE)

Energy-aware Routing to a Mobile Gateway (EARMG) Most of the energy aware protocols proposed consider the BS to be stationary which may not be true for all applications. Thus some modifications over the existing protocols have to be suggested so as to consider the mobility of the BS. EARMG is one such algorithm [31]. Location of the gateway has to be intimated to other nodes so as to re discover routes.

SIDE is defined for large number of stationary sink nodes [36]. When a set of nodes in the target region receives a query message, one of them will be chosen as a source and it does the data fusion.If a data has to be sent to multiple sinks, instead of sending the data to all sinks via separate paths, source sends the data to one of the sinks and the sink nodes will then share the information amongst themselves.

Hybrid Energy-Efficient Distributed clustering (HEED)

Improved Weighting Clustering Algorithm (IWCA) IWCA is a cluster-based algorithm in which a node

HEED is a dynamic clustering algorithm where the



931


Broadcast (IB) packet to its neighbors. Each of the nodes receiving the IB packet, extracts the table information and stores in it and updates the fields like hop count and cost and forwards further [42]. The IB packets will be forwarded in the network repeatedly till all nodes update its Hop count properly. Depending on the remaining energy in the node, the cost factor can be either linear, quadratic or cubic.

with higher degree is selected as the CH [37]. It is more applicable for mobile nodes. A node with less mobility is chosen as a CH. CH selection is based on its residual energy, mobility and its distance to its neighbors Genetic Annealing based Clustering Algorithm (GACA) GACA is hierarchical routing protocol in which the CHs are selected on the iterative basis [37]. Genetic Algorithm (GA) or Simulated Annealing (SA) is considered to select the optimal CHs.

Minimum Transmission Hierarchy (MTECH)

Energy

with

Clustering

MTECH is a hierarchical routing protocol that uses cluster model. Node in the cluster having the highest energy will be chosen as a CH [43].

Base-Station Controlled Dynamic Clustering Protocol (BCDCP) In [38] authors have proposed a centralized cluster based routing. Initially BS receives the energy levels of all nodes and computes a set of nodes with energy more than the average energy. Nodes from that set will be chosen as CHs. Cluster formation is done in iterative manner starting with two clusters in the network. Nodes will be allocated to respective clusters depending on the distance. BS will then computes the minimum paths using spanning tree approach and will be intimated to the nodes.

Energy Efficient Clustering Routing Algorithm (EECR) One of the ways to incorporate the energy efficiency is to design a protocol so as to distribute the load uniformly amongst nodes. EECR is a one such cluster based algorithm [44]. BS does cluster formation and the selection of CHs. Algorithm is divided into two phases cluster formation and data transmission. BS does the job of cluster formation by iteratively dividing the network into sub networks. Reliable Energy Aware Routing (REAR)

Cluster based Energy Efficient Routing Protocol (CBEERP)

REAR is an on demand routing protocol in which the energy extinction is reduced by avoiding retransmission of the packets [45]. Algorithm works in four phases: Path discovery, energy reservation, reliable transmission and reserved resource release. When a node receives an interest message, it checks for a path to the BS. If path does not exists, service path discovery is initiated. In the meanwhile BS will discover a backup path to the sender. Energy is reserved in the path depending on the requirement.

CBEERP is a cluster based routing protocol without considering the location information of the nodes. The algorithm involves two phases: cluster construction phase and data transmission phase. Initially BS broadcasts an advertisement message for CH selection. Once CHs are chosen, they advertise to other nodes and all nodes will join appropriate clusters [39], [24]. Optimal Energy-Efficient Routing (OEER) It is a table driven routing protocol in which the Bellman Ford Algorithm is incorporated for routing. OEER balances the minimum and average node lifetime [40]. Routing problem is formulated as a non-linear optimization problem. Langrangean Relaxation is being applied to solve the problem

Energy Efficient Clustering Algorithm (EECA) EECCA is a centralized clustering algorithm in which the whole network topology will be notified to the BS using notification algorithm. The BS then will decide the clusters. Notification protocol involves two phases: ascending phase from nodes to the BS telling their existence and the descending phase from the BS to the nodes informing the cluster to which it belongs [46].

Energy Efficient AODV (EEAODV) An improvement over AODV has been proposed in which the residual energy of the intermediate nodes will be considered [41]. This has been accomplished by introducing an additional field in the RREQ packet, Minimum Residual Energy (Min-RE). Each of the intermediate nodes updates the Min- RE field. After destination receiving the RREQ messages, the ratio of Min RE and hop count is calculated and the path with largest ratio is chosen.

Distributed (DEEC)

Energy-Efficient

Clustering

Algorithm

DEEC is a distributed algorithm for heterogeneous network [47]. The CHs are elected based on the ratio of the residual energy and the average energy in the network. Nodes with more initial and residual energy have more chances of becoming the CH.

Energy Aware Distance Vector Routing Protocol (EADV)

Energy-Efficient Clustering Algorithm (ECA) Energy efficient Clustering Algorithm (ECA) is a dynamic clustering algorithm [48]. Algorithm involves

In EADV, initially the sink node broadcasts Initial



932


CH.

two phases viz setup phase and steady state phase. Time stamp and TTL (Time to Live) is used in the message to adjust the diameter of the cluster. Distance-Based (DPDD)

Proportional

Delay

Minimum Energy Relay Routing (MERR) MERR takes the linear topology of the network into consideration. Here it is assumed that the sensor nodes transmit the data to the base stations via relay nodes and the distance between the relay nodes has to be approximately equal to the characteristic distance [54]. Characteristic distance is the optimal distance which is a constant predetermined at the time of network set up.

Differentiation

Most of the energy efficient routing algorithms do not offer a good QoS which are not suitable for real time applications. DPDD is a QoS assured energy efficient routing in which end to end delay requirement is satisfied [49]. It is assumed that all nodes will be aware of their distance from the BS. A parameter r is defined to allocate the BW for real time and non real time traffic.

Energy Aware Routing (ERP) Energy aware routing uses energy availability and the received signal strength from the nodes to determine the optimal path. Each of the nodes will decide the next node to which the data has to be forwarded based on its residual energy and the signal strength. The algorithm operates in three phases: Neighbor Discovery, Route Reply and Reliable Transmission [55], [56]

Maximum Energy Cluster Head (MECH) Although LEACH was proved to be one of the optimal protocols, it does not consider the node distribution. MECH is an improvement over LEACH. It involves three phases: Setup, steady and forward phases [50]. Initially every node will broadcast hello packets to its immediate neighbors. If the number of neighboring nodes reaches a predetermined value NH, the corresponding node becomes the CH and all its immediate neighbors will become the members. After backoff time every node will reselect the CH depending on the signal strength.

Transmission Power Control MAC Protocol (SMAC) Energy consumption can be reduced by reducing the idle time of the sensor nodes. In SMAC, nodes form virtual clusters based on common sleep schedules to reduce control overhead and enable traffic-adaptive wakeup [57].

Hop-based Energy Aware Routing Algorithm (HEAR)

Cross Layer MAC (CLMAC) Protocol

HEAR algorithm does not consider data combining and routing overhead reduction [51]. Initially the BS will collect the information about all the sensor nodes in WSN. If a node has any information to be communicated to the BS, a message will be sent to the BS. Depending on the distance, BS determines an optimal hop count. It also determines the corresponding hops to reach the node.

Reducing the size of the routing table will minimize the consumption of the energy. CLMAC protocol includes routing distance in the preamble field [58]. As the usage of big routing tables has been replaced by a field in the preamble, the traffic in the network can be reduced thus reducing the energy consumption. Energy Efficient Cluster Head Selection Algorithm (EECSA)

Energy Efficient Adaptive Multipath Routing (EEAMR) Multipath routing is one possible way of achieving energy awareness as it leads to the distribution of load along multiple paths to the destination. But finding the optimal number of paths is a vital issue. EEAMR is a low overhead multipath routing with energy awareness [52]. Generally node with highest energy and farthest from the transmission node is selected as the next hop. It involves two phases: Multipath construction phase and data transmission phase.

EECSA is a cluster based routing algorithm [59]. The algorithm works in three phases: CH selection, cluster formation and the scheduling based on TDMA. CHs are selected based on the residual energy. Proposed algorithm is an improvement over LEACH. If the available energy of the node is greater than the 50% of the initial value then the normal LEACH is used else the proposed protocol is used in which the probability of the node to be selected as a CH is the ratio of the residual energy and the initial energy.

Energy Efficient Clustering Scheme (EECS)

Simple Energy Efficient Routing Protocol (SEER)

EECS is a cluster-based approach developed for periodical data gathering applications [53]. It focuses on low control overhead and uniform load distribution. Initially BS broadcasts a HELLO message from which all nodes will compute their distance from the BS. In CH formation phase, nodes that are interested to become CHs will advertise a message within its radio range. A node with higher energy level is elected as the CH for the corresponding cluster. Other members of the cluster will be decided based on the distance from the

In [60] a flat algorithm has been proposed to improve the network lifetime. SEER reduces the overall traffic in the network thus decreasing the energy consumption. In the initialization phase, BS will broadcast the packet. All nodes receiving it will update with hop count and re broadcast to its neighbors by replacing the source address by itself and enters its residual energy. For forwarding node selects its neighbor with the hop count less than itself. If more nodes have the same minimum hop count, the one with



933


sink, which leads to unbalanced energy distribution. In EADD protocol, if two sources receive the same interest message, both the nodes will respond to the destination via different paths [67]. The path with the maximum residual energy is chosen for communication. If more than one path are having the same available total energy, then the path involving the highest minimum energy will be selected.

maximum energy is selected. Same process is repeated at the intermediate nodes. Energy-Efficient Multipath Routing Protocol (EEMR) EEMR is multipath routing [61]. When there will be an event, the surrounding nodes exchange data themselves and one of them is chosen as the source. Source aggregates the data and sends to the BS. Each node selects its next hop depending on the distance and the residual energy. It is assumed that multipaths are disjoint. EEMR involves four phases: Initialization, multipath selection, data transmission and path maintenance.

Energy Efficient Routing Scheme for Mobile Wireless Sensor Networks (MLEACH) Its an improvement over existing LEACH algorithm which incorporates the mobility of the sensors. It is assumed that all nodes are homogeneous and position aware and the BS is stationary [68]. The total sensing area is divided into sub areas and CH is optimized within the sub areas. Generally nodes with less mobility are preferred to the CH.

Energy Efficient Clustering Algorithm (EECA) In EECA, CHs are distributed evenly in the network and unnecessary CHs are avoided [62]. Clusters have to be formed in the network in such a way that there will an uniform distribution of the CHs.Advertisement message can be broadcasted based on CSMA protocol. Once the CHs are elected, clusters will be formed.

Color-theory-based Energy Efficient Routing (CEER) CEER is a cluster based algorithm based on color theory based localized algorithm. Location of a mobile node is represented by its RGB values. Server computes the positions of the nodes depending on the RGB information. When a node moves to a new location, it collects the RGB information from its neighbors. CEER involves three phases viz setup phase, data dissemination phase and refinement phase [69].

Reactive Energy Decision Routing Protocol (REDRP) REDRP is a reactive energy aware protocol [63]. It involves four phases viz Initialization, route discovery, data transmission and route maintenance.Initially the BS broadcasts a timer packet, each node records the time stamp at the distance field as less is the delay near is the node to the BS. Route is established only when there is an event.

Extending Lifetime of CH (ELCH) ELCH is a hybrid type protocol in which the direct communication is used for any intra cluster transmission and multihop is used for inter cluster communication. It uses MTE (Minimum Transmission Energy) protocol as the underlying protocol. It involves two phases: setup phase and steady state phase [70].

SeNsOr netWork CLUSTERing (SNOW) SNOW is a cluster-based algorithm in which the nodes with higher residual energy are chosen to be the CHs. After CHs are formed, the BS selects the region heads (RH) amongst the CHs.Nodes with higher residual energy will be chosen as the region heads. After receiving the intimation from the BS, each of the CHs check whether it is a RH and accordingly it will set its region ID. CHs and RHs are chosen so as to distribute the load amongst them [64].

Hybrid Energy Aware Routing Protocol (HEARP) In [71] a new protocol was proposed which combines the features of LEACH and PEGASIS. In HEARP members in the clusters will not communicate directly with the CHs but through the intermediate nodes. It involves two phases: initialization or set up phase and the steady state phase

Energy Efficient Dynamic Clustering (EEDC) EEPA is an energy efficient protocol based on the metric [65]. Initially nodes with higher energy form the CHs and the remaining nodes become the members. Clusters are formed so as to minimize the distance between the cluster members and the CH. Cluster updating is done in the same manner as that of LEACH [24].

Global Simulated (GSAGA)

Annealing

Genetic

Algorithm

GSAGA is a centralized control algorithm which involves two phases: Setup phase and steady state phase [72]. Initially all nodes will transmit the location information and the residual energy to the BS. BS will then compute the average network energy. Nodes with energy more than the average can become the CHs. Genetic algorithm is used for the cluster head selection

Energy Efficient Geographic Grid Routing (EEGGR) EEGGR considers the BS to be mobile. Sensor nodes are considered to have location awareness. Grid structure is used for forwarding the data to the sink node [66].

Advanced Medium Access Control (A-MAC)

Energy Aware Directed Diffusion (EADD) Protocol

A-MAC is a TDMA-based MAC protocol, which uses a distributed technique where node selects its own slot by collecting the information from its neighbors [73]. A-

The normal Directed Diffusion algorithm will always considers the shortest path between the source and the



934


assumed that BS can predict its movement for the next time interval. The network topology can be computed by the BS by predicting its and the nodes movement profile. This topology remains valid for the time interval under consideration.

MAC has four states in its operation namely initial, wait, discover, and active states Track-Sector Clustering (TSC) TSC is a cluster-based algorithm in which the network is divided into concentric circular tracks and triangular sectors [74]. BS does the formation of the tracks and clusters. TSC uses tracks and sectors to form clusters. Head nodes in each track are selected by the BS. A node is selected randomly as the head node in level 1. Nodes with the similar slopes will be chosen as the head nodes in the higher levels.

Improvement on LEACH Protocol (VLEACH) Many improvements over LEACH have been proposed. VLEACH is one such protocol [82]. In VLEACH there will be a CH, a vice CH and cluster members. Vice CH will become the CH if the current CH dies. Provisioning of the vice CH is the key parameter in VLEACH.

Partition LEACH Algorithm (PLEACH)

Energy Efficient Heterogeneous Clustering (EEHC)

In [75] an improvement over the LEACH algorithm was proposed: PLEACH. It first does the optimal partitioning of the network and then the node with the highest energy in each partition will be chosen as the CH. It outperforms LEACH as the CHs are evenly distributed over the network [76].

Cluster based routing is more advantageous compared to flat routing. EEHC is a heterogeneous cluster based routing. We can categorize the nodes present in the network as normal nodes with limited energy and advanced and super nodes with higher energy. Weighted probability is considered for the election of the CHs [83].

Energy Aware Adaptive Clustering (EAAC)

Energy Efficient Routing Algorithm for Hierarchically Clustering (ERHC)

Cluster based algorithms were proved to be better compared to flat routing. EAAC is one such algorithm in which CHs and the next heads are determined based on the residual energy and the distance between the CH and the members in the cluster [77]. EAAC protocol works in various rounds each has a set up phase and a steady state phase.

ERHC is a cluster-based algorithm. Hop count from the BS is considered to form the hierarchy and the CHs are selected in an autonomous manner. Alternative sensor node for all intermediate nodes is determined in this algorithm where the determined node will become the next alternative intermediate hop if the energy of the present hop goes below the threshold [84].

Energy-Level Passive Clustering (ELPC)

Energy Aware DSR (EADSR)

ELPC is a passive clustering algorithm in which clusters are formed on demand. It focuses on two issues viz: minimizing the energy per packet and uniform load balancing [78].

Traditional DSR can be slightly modified to incorporate the energy function. The basic idea behind this is as follows: when an intermediate node in the network decides to forward a RREQ message that it has received, it introduces an additional delay before retransmitting this message. The delay is dependent on the residual battery power in the node. Thus the nodes with higher battery levels are more likely to be included in the path [85].

MiSense Hierarchical Cluster-Based Routing Algorithm (MiCRA) MiCRA is an extension over the HEED protocol [79]. It involves two levels of cluster hierarchy. First level CHs elects the second level CHs. MiCRA considers two parameters viz residual energy of the nodes which is used to select the CHs and the intra cluster communication costs used to break ties.

Energy Efficient Clustering Hierarchy and Data Accumulation (EECHDA) In the protocol proposed in [86], cluster head performs the communication with the base station. It involves two phases viz Cluster head election phase and data transfer phase. After some time slots a non CH with higher energy becomes the new CH.

GRAdient Cost Establishment (GRACE) GRACE is a dynamic routing algorithm in which BS initializes the set up phase by sending an advertisement packet. Routing table is updated at all nodes depending on the energy and distance parameters. Each of the nodes will then forward the packet from the source node to the node with the minimum cost [80]

An Adaptive Energy Efficient Reliable Routing Protocol (AEERRP) In the AEERRP, the source adjusts the flooding rate depending on the loss rate at the sink [87]. If the loss rate is very less, then the transmission power can be reduced. Thus there has to be a tradeoff between the power consumption and the latency.

Sensor system for Hierarchical Information gathering through Virtual triangular Areas (SHIVA) In SHIVA both nodes and BS are mobile [81]. Although the nodes are mobile, it is assumed that the logical cluster remains same for certain duration. It is



935


It is an improvement over SEER [60]. A flag is used in the packet to distinguish the normal and the critical data.

Energy-Aware QoS Routing (EAQOS) Although the proposed energy aware routing protocols minimizes the energy consumption, they do not work good for some applications where QoS is also required. EAQoS is an energy aware routing protocol, which also assures good QoS [88]. It is a cluster-based protocol where the cluster formation is done by the command node. In order to support both real time and non real time traffic, a ratio r is defined is the initial value set by the gateway and represents the amount of bandwidth to be dedicated both to the real-time and non-real-time traffic on a particular outgoing link in case of a congestion.

Energy Efficient Routing Protocol (EERP) EERP is based on the learning automata [94]. It does efficient flooding which in turn leads to energy efficient routing. Protocol has two phases: identification phase and data transmission phase. Threshold Distributed Energy Efficient Clustering (TDEEC) Although cluster based routing is found to be better selection of the cluster head is a crucial issue. TDEEC is a hierarchical routing in which CH selection is based on the residual energy of the node and the average energy of the network [95].

Energy Efficient Cluster-based Routing Algorithm (EECRA) EECRA is a cluster based routing algorithm, which assumes that the sensors are deployed randomly over the given geographical area [89]. CH is selected based on two parameters: residual energy and the node degree. CH selects the members for the cluster based on the energy and the distance from itself.

Clustering Technique for Wireless Sensor Networks (CTRWSN) In [96] a cluster based algorithm was proposed in which the CHs are chosen on the rotation based in order to have uniform energy depletion. It minimizes the energy consumption for new CH selection after each round by keeping the selected CH for m consecutive rounds. Two level heterogeneous network with normal and advanced nodes is considered. There are two phases of operation: Setup phase and steady state phase.

Homogenous Clustering Algorithm (HCA) For cluster based routing, cluster formation and leader election are two crucial issues. In homogeneous clustering sensors will be of same hardware and initial battery capability. Initially BS collects the information about the location of all nodes and initializes clusters such that all cluster head selection is uniform throughout the area. Initially CHs are selected randomly in each zone. New CH is elected periodically depending on the residual energy and the relative distance from the current CH [90].

Location Aware Multi-level Clustering (LAMC) Multi level clustering algorithms were developed in the literature. An improvement over EEMC is proposed in [97]. In LAMC, BS will broadcast a beacon message and all nodes will reply with their location and the residual energy. BS will then send the command message and the CHs for level 1 are selected. CHs of level 1 will broadcast message to all nodes within a certain range and the process is repeated within the cluster to select level 2 CHs. This process is continued for a predefined number of times to have multi level clustering.

Energy Efficient Cross Layer Routing Algorithm with Dynamic Retransmission for Wireless Sensor Networks (E2XLRADR) In E2XLRADR a cross layer approach is considered which involves sharing of information amongst layers. Physical, MAC and network layers are considered. Algorithm involves five phases [91]. Ant colony optimization is combined with Opportunistic Routing Entropy (ACO-TDOP)

Power Aware Multi-level Clustering (PAMC) In PAMC nodes need not have to have their location information [97]. BS will broadcast a beacon message. All nodes reply with the minimum power level to reach the BS along with the residual energy and the power information. Further process is same as that of the LAMC algorithm [97].

the

Opportunistic Routing is an effective energy aware routing in which the next relay is selected dynamically for each hop and packet. Key design parameter includes the selection of the path with minimum delay and higher energy level. For better performance one has to choose the next hop which has more energy level, consumes less energy and is nearby to the sink [92].

Distance based Energy Aware Routing Algorithm (DEAR) Goal of the DEAR algorithm is to optimize all individual paths so as to make all nodes to consume energy at the same rate [98]. It consists of two phases viz route setup and route maintenance. Whenever there is a data to be transmitted, source node initiates the route setup phase.

Balanced Energy-Aware Routing (BEAR) Energy aware routing protocols are classified as either energy saver or energy manager. In BEAR there will be a trade off between energy balancing and optimal distance [93]. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


936


Tree based Energy and Congestion Aware Routing Protocol (TECARP)

Cluster Based Energy Efficient Routing Protocol (CBERP)

Congestion avoidance is more efficient than congestion control as congestion control requires more resources. TECARP is a hierarchical routing protocol which considers the energy awareness through congestion control [99]. It focuses on congestion avoidance by constructing the tree efficiently. It involves three phases: network clustering, creating routing tree and data forwarding.

Although LEACH is an efficient algorithm proposed for WSNs, CHs will die earlier than other nodes [24]. CBERP is an improvement over LEACH. BS selects the CHs initially. Multihop transmission using chain of CHs is done in CBERP [102]. Optimal Path Energy Efficient Routing (OPEER) Although EEAODV [41] was proved to perform better, it has been still improved by assigning the job of route establishment to the BS in OPEER [103, 104]. As the BS is not energy constrained, it can be over burdened without affecting the network performance. Also the usage of multiple thresholds has been proposed in the paper, which further assures the uniform load balancing in the network.

Energy Efficient Grid Clustering (EEGC) EEGC is a cluster based energy aware routing protocol. It involves two basic principles: path with minimal energy consumption and load balancing. EEGC normalizes the clustering area. It overcomes the drawback of uneven area distribution for the clusters in the LEACH algorithm [100]. A Tree Based Routing Protocol (TBRP)

IV.

TBRP is an energy aware routing protocol proposed for mobile sensor networks. In this protocol all nodes in the network form a tree with different levels. Distance between levels is equal to the radio communication range. The algorithm involves three phases: tree formation, data collection and transmission and purification phases [101].

Comparative Analysis

In this section we will be presenting a qualitative analysis of the protocols discussed in Section III. The comparison has been made considering various network parameters like the delay, routing overhead, QoS, type of routing protocol etc.

TABLE I COMPARATIVE ANALYSIS OF THE ENERGY AWARE ROUTING PROTOCOLS Query Data Scalability Overhead Position Protocol Classification Based Delay Mobility Aggregation Estimation LEACH Hierarchical/ [24] Cluster based

Yes

Limited

No

Relative positions are Stationary considered Depends on Depends on Relative Depends on the the positions are the underlying underlying considered underlying protocol protocol protocol Less

Less

Working Layer

QoS

Network

Good

MAC/ Network

Good

GAF [25]

Flat

No

Good

NA

GEAR [26]

Flat

No

Good

No

Less

Slightly high Considered

Stationary

Network

Limited

TEEN [27]

Hierarchical/ Cluster based

Yes

Limited

Yes (Reactive)

Less

Relative Depends on positions are Stationary the threshold considered

Network

Good

EER [28]

Flat

Yes

Yes

No

Less

Stationary

Network

Good

Relative Slightly high positions are Stationary considered

Network

Limited

Stationary

Network

Limited

Mobile

Network

Limited

Quasi Stationary

Network

Good

No

Stationary

MAC/ Network

Good

No

Stationary

Network

Good

Relative positions are Stationary considered

MAC and Network

Limited

Network

Limited

PEGASIS Hierarchical/ [29] Cluster based EBRP Flat [30] EARMG Flat [31] HEED Hierarchical/ [32] Cluster based GSP Flat [33] EARP Flat [34] RAW-E [35] CODE [36]

Flat

Flat

Less

No

Yes

Limited

No

Less

Yes

Good

Yes

Slightly high

NA

Good

Yes

Slightly high Slightly high Considered

Yes

Limited

No

No

Very Good

No

Very Less

Less

No

No

Yes

Less

Less

No

Limited

No

No

Good

Yes

Less

Less

Less

More

Less

Considered

Considered



Mobile


937


Protocol Classification

Query Data Scalability Based Aggregation

SIDE [36]

Flat

Yes

Good

Yes

IWCA [37]


Yes

Limited

No

GACA [37]


Yes

Limited

No

BCDCP Hierarchical/ [38] Cluster based

Yes

Limited

No

CBEERP Hierarchical/ [39] Cluster based

Yes

Limited

No

No

Limited

No

No

Yes

Yes

OEER Flat [40] EE AODV Reactive/ Flat [41] EADV Flat [42] Hierarchical/ MTECH Cluster based [43] EECRA Hierarchical/ [44] Cluster based REAR Reactive [45] EECCA Hierarchical/ [46] Cluster based

Overhead

Position Estimation

Mobility

Working Layer

QoS

Slightly high Considered

Stationary

Network

Limited

Network

Limited

Network

Limited

Network/ MAC

Limited

Network

Good

Network

Good

Network

Limited

Network

Limited

Network

Limited

Network

Limited

Delay

Less

Relative Slightly high positions are Mobile considered Relative Slightly high Less positions are Mobile considered Relative Slightly high Less positions are Stationary considered Relative Less Less positions are Stationary considered Stationary Less Less No Less

More

More

No

Stationary

Can be Dynamic More Can be high considered network Relative Can be Less Less positions are mobile considered Relative Slightly high Less positions are Stationary considered

No

Yes

Yes

Yes

Limited

Yes

Yes

Good

No

No

Good

Yes

Slightly high Slightly high

Yes

Limited

Yes


No

Stationary Network/transport

Limited

Stationary

Network

Limited

Network

Limited

Network

Good

Network

Very Good

MAC and Network

Limited

Network

Good

Network

Good

Relative positions are Stationary considered Relative positions are Stationary considered

DEEC [47]


Yes

Limited

No

Slightly high

Less

ECA [48]


Yes

Limited

No

Less

Less

DPDD [49]


Yes

Limited

No

Slightly high

MECH [50]


Yes

Limited

No

More

No

Yes

No

NA

Can be high

No

Yes

Good

No

Less

Less

No

Yes

Limited

No

Slightly high

Less

Considered

Stationary

Network

Limited

No

Limited

Yes

Less

Less

No

Stationary

Network

Depend on dmean

Flat

No

Yes

No

Less

More


Network

Limited

Flat

No

Yes

No

More

Less

No

Stationary

MAC

Good

No

Yes

No

Very less

Less

No

Stationary MAC & Network

Yes

Limited

No

Less

Less

Flat

No

Good

No

Slightly high

Less

Flat/multipath

Yes

Good

No

Less

Less

EECA [62]


Yes

Limited

No

Less

Less

REDRP [63]

Reactive

No

Good

Yes


SNOW [64]


Yes

Limited

No

Slightly high

HEAR Flat [51] EEAMR Flat/multipath [52] EECS Hierarchical/ [53] Cluster based MERR Flat [54] ERP [55, 56] SMAC [57] CLMAC [58]

Flat

EECSA Hierarchical/ [59] Cluster based SEER [60] EEMR [61]

Less for real Considered Stationary time traffic Relative Not Slightly high positions are considered considered


Less

Stationary Stationary

Relative positions are Stationary considered Stationary No

Limited

Network

Good

Network

Limited

Stationary

Network

Good


MAC/ Network

Good

No

No

Stationary

Network

Limited

Considered

Stationary

Network

Limited


938


Protocol Classification EEPA [65] EEGGR [66] EADD [67]


Overhead

Delay


Yes

Limited

No

Slightly high

Flat

Yes

Good

Yes

Less

No

Yes

Yes

More

Flat

MLEACH Hierarchical/ [68] Cluster based

Yes

Limited

No

Less

Position Estimation

Mobility

Working Layer

QoS


Network/ MAC

Limited

Mobile

Network

Limited

No

Stationary

Network

Limited

Considered

Nodes are Mobile but not BS

Network

Good

Mobile

Network

Limited

Network

Limited

Network

Limited

Network

Limited

Slightly high Considered More

Less

Less

CEER [69]


Yes

Limited

No

ELCH [70]


Yes

Limited

No

HEARP Hierarchical/ [71] Cluster based

Yes

Limited

No

GSAGA Centralized/ [72] Cluster based

Yes

Limited

Yes

NA

NA

Yes

NA

Less

NA

Stationary

MAC

Good

Yes

Limited

Yes

Slightly high

Less

Considered

Stationary

Network

Limited

Yes

Limited

Yes

Slightly high

Less

Considered

Stationary

Network

Limited

Less

Considered

Stationary

Network

Good

Network

Good

Network

Limited

A-MAC [73] TSC [74] PLEACH [75, 76] EAAC [77]

Data centric Centralized/ Cluster based Centralized/ Cluster based Centralized/ Cluster based

ELPC [78]


Considered

Relative positions are Stationary considered Relative Less Slightly high positions are Stationary considered Relative Slightly high Less positions are Stationary considered

Slightly high

Less

Less

Yes

Limited

Yes


Yes

Limited

No

Less

Less

MiCRA Hierarchical/ [79] Cluster based

Yes

Limited

No

Slightly high

Less

No

Good

Yes

Less

Slightly high

No

Stationary

Network

Limited

Yes

Good

No

Less

Less

Considered

Mobile

Network

Good

Yes

Limited

No

Less

Less

Network

Good

Network

Limited

Network

Good

Stationary

Network

Limited

Relative Not positions are considered considered

Network

Limited

GRACE Flat [80] SHIVA Hierarchical/ [81] Cluster based VLEACH Hierarchical/ [82] Cluster based EEHC [83]


Yes

Limited

No

ERHC [84]


Yes

Limited

Yes

Less

Less

EADSR [85]

Flat

No

Yes

No

Less

Slightly high

EECHDA Hierarchical/ [86] Cluster based

Yes

Limited

No

Less

Less

AEERRP [87]

No

Yes

Yes

Less

Less

Flat

EAQOS Hierarchical/ [88] Cluster based

Yes

Limited

No

EECR [89]


Yes

Limited

No

HCA [90]


Yes

Limited

No

Flat

No

Good

No

Flat

No

Good

No

Flat

No

Good

No

E2XLRA [91] ACO TDOP [92] BEAR [93]

Slightly high

Less

Relative positions are Stationary considered Relative positions are Stationary considered

Relative positions are Stationary considered Relative positions are Stationary considered Relative positions are Stationary considered No

No

Stationary

Depends on Relative Less design positions are Stationary parameter r considered Relative Less Less positions are Stationary considered BS will have Slightly less Less all location Stationary information Less Slightly high

Less


Network

Limited

Network

Depend on design parameter

Network

Good

Network

Good

Less

No

Stationary

Physical, MAC & Network

Good

Less

No

Stationary

Network

Limited

Less

No

Network

Good

Stationary


939


Protocol Classification EERP [94]

Overhead

Delay

Position Estimation

Less

No

No

Good

No

Less

TDEEC Hierarchical/ [95] Cluster based

Yes

Limited

No

Slightly high

Hierarchical/ CTRWSN Cluster based [96]

Yes

Limited

No

Less

Less

Yes

Limited

No

High

Less

Yes

Limited

No

High

Less

No

Good

No

Slightly high

Less

LAMC [97] PAMC [97] DEAR [98]

Flat


Hierarchical/ Cluster based Hierarchical/ Cluster based Flat

TECARP Hierarchical/ [99] Cluster based

Slightly high

Yes

Limited

No

Slightly high

Less

EEGC [100]


Yes

Limited

No

Slightly high

Less

TBRP [101]

Hierarchical

Yes

Limited

Yes

Slightly high

Less

CBERP Hierarchical/ [102] Cluster based

Yes

Limited

Yes

Slightly high

Less

OPEER [103]

No

Good

Yes

Very Less

Less

Flat/ Centralized

[3]

Table I summarizes the behaviour of each of the protocols studied. A detailed summary has been presented in Table I. It has been observed from the table that an ample number of protocols were proposed in the literature [105], [106], [107], [108], [109]. Most of the algorithms were application specific and may not work fine for all type of application environments. Most of the proposed energy aware protocols trade off delay and energy efficiency.

[4] [5] [6] [7] [8] [9]

V.

Conclusion

[10]

Routing in WSNs is one of the emerging areas of research. There are many challenging tasks to be considered while proposing a routing protocol for WSN. Energy awareness is one of the important parameters the routing protocol should posses. In this paper a comprehensive survey has been made on the energy efficient routing protocols and an analysis has been presented. The comparison has been made based on various parameters. From the table it is clear that, there is no clear winner. The selection of the routing protocol has to be made based on the application. But we can conclude that cluster based algorithms are better compared to flat routing.

[11] [12]

[13]

[14]

[15]

[16]

References [1] [2]

David Culler, Deborah Estrin, Mani Srivastava, Overview of Sensor Networks, IEEE Computer Society, August 2004 C Siva Ram Murthy and B S Manoj, Adhoc Wireless NetworksArchitectures and Protocol ( Pearson education, 2004)

[17]


Mobility

Working Layer

QoS

Network

Good

Network

Limited

Network

Good

Stationary

Network

Limited


Network

Limited

Network

Limited

Network

Good

Network

Limited

Network/ MAC

Limited

Network

Limited

Network

Good

Stationary

Relative Stationary positions are considered Relative positions are Stationary considered Yes

No

Stationary

Relative positions are Stationary considered Relative positions are Stationary considered Relative positions are Mobile considered Relative positions are Stationary considered No

Stationary

Andrew S Tanenbaum, Computer Networks (4e, Pearson Education, 2003) Behrouz A Fouruuzan, Data Communications and Networking, (3e, McGrawHill Publication, 2004) Theodore. S Rappaport, Wireless Communications – Principles & Practices (Pearson Education, 2nd Edition, 2002). Mullet, Introduction to Wireless Telecommunication Systems and Networks (Cengage Learning Publication, 2006) Kamilo Feher, Wireless Digital Communications (PHI Publication) Mohammed Iliyas, A Handbook on Wireless Adhoc Networks (CRC Press, 2003) Vijay Garg, Wireless Communication and Networking, Elsevier India, 2007 Michel Daoud Yacoub, Wireless Technology: Protocols, Standards, and Techniques (CRC Press, 2002) Ivan Stojmenvoic, Handbook of Wireless Networking and Mobile Computing (John Wiley and Sons, 2002) Curt Schurgers, Mani B. Srivastava, Energy Efficient Routing in Wireless Sensor Networks, http://www.citeseerx.ist.psu.edu/viewdoc/ download?doi=10.1.1.11 Sinem Coleri Ergen and Pravin Varaiya, On Multi-Hop Routing for Energy Efficiency, IEEE Communications Letters, Vol. 9, No. 10, October 2005 M.K.Jeya Kumar, Evaluation of Energy-Aware QoS Routing Protocol for Ad Hoc Wireless Sensor Networks, International Journal of Electrical, Computer, and Systems Engineering 4:3 2010 Rajashree.V.Biradar, V.C .Patil, S. R. Sawant, R. R. Mudholkar, Classification and Comparison of Routing Protocols in Wireless Sensor Networks, Special Issue on Ubiquitous Computing Security Systems, UbiCC Journal – Volume 4, pp. 704-711, 2009 T. Ezzedine, M. Miladi, R. Bouallegue, A Performance Comparison of Energy Efficient MAC Protocols for Wireless Sensor Networks, International Review on Computers and Software (IRECOS), Vol 3, No 4, July 2008. Kemal Akkaya , Mohamed Younis, A Survey on Routing Protocols for Wireless Sensor Networks, Elsevier, Ad Hoc Networks 3 (2005) 325–349


940


[18] Jaspal Kumar, Muralidhar Kulkarni, Daya Gupta, Performance Comparison of MANET Routing Protocols, International Review on Computers and Software (IRECOS), Vol 5, No 1, Jan 2010. [19] D B Johnson and D A Maltz, DSR-The Dynamic Source Routing Protocol for Multihop Wireless Adhoc Networking Addison Wesley,2001 [20] M C Domingo, D Remondo and O. Leon, A Simple Routing Scheme for Improving Adhoc Network Survivability, GLOBECOM, IEEE, 2003 [21] Maleq Khan, Gopal Pandurangan, Energy Efficient Routing Schemes for Sensor Networks, Research Thesis, Department of Computer Science, Purdue University, West Lafayette, 2003 [22] Raihan Ahmed, A Survey on the Routing Protocols in Wireless Sensor Networks, MTech Thesis, Department of Electrical and Electronic Engineering, BRAC University, Dhaka, Bangladesh, August 2010 [23] Chih-Wei Hsiao, Optimal Energy-Efficient Routing for Wireless Sensor Networks, Master Thesis, Department of Information Management, National Taiwan University, Taiwan, 2004 [24] Wendi Rabiner Heinzelman, Anantha Chandrakasan, and Hari Balakrishnan, Energy-Efficient Communication Protocol for Wireless Microsensor Networks, Proceedings of the 33rd Hawaii International Conference on System Sciences, IEEE 2000 [25] Y. Xu, J. Heidemannn, and D. Estrin. Geography-Informed Energy Conservation for Ad Hoc Routing. In Proc. of the Seventh Annual ACM/IEEE International Conference on Mobile Computing and Networking (MobiCom 2001), Rome, Italy, July 2001. [26] Yan Yu, Ramesh Govindan, Deborah Estrin, Geographical and Energy Aware Routing: A Recursive Data Dissemination Protocol for Wireless Sensor Networks, UCLA Computer Science Department Technical Report UCLA/CSD-TR-01-0023, May 2001 [27] A. Manjeshwar, D.P. Agrawal, TEEN: a Protocol for Enhanced Efficiency in Wireless Sensor Networks, Proceedings of the 1st International Workshop on Parallel and Distributed Computing Issues in Wireless Networks and Mobile Computing, San Francisco, CA, IEEE April 2001 [28] Curt Schurgers, Mani B. Srivastava, Energy Efficient Routing in Wireless Sensor Networks, http://sciencestage.com/d/1613070/energy-efficient-routing-inwireless-sensor-networks-2001-.html [29] S. Lindsey, C. Raghavendra, PEGASIS: Power-Efficient Gathering in Sensor Information Systems, IEEE Aerospace Conference Proceedings, 2002, Vol. 3, 9-16 pp. 1125-1130 [30] Sasanka Madiraju, Cariappa Mallanda, Rajgopal Kannan, Arjan Durresi and S.S.Iyengar, EBRP: Energy Band based Routing Protocol for Wireless Sensor Networks, Proceedings of the International Conference on Intelligent Sensors, Sensor Networks and Information Processing, IEEE 2004 [31] Kemal Akkaya and Mohamed Younis, Energy-aware Routing to a Mobile Gateway in Wireless Sensor Networks, International Conference on Global Telecommunication, IEEE 2004 [32] Ossama Younis and Sonia Fahmy, HEED: A Hybrid, EnergyEfficient, Distributed Clustering Approach for Ad Hoc Sensor Networks, IEEE Transactions on Mobile Computing, Vol. 3, No. 4, October-December 2004 [33] Xiaobing Hou and Debdhanit Yupho, Gossip Based Sleep Protocol for Energy Efficient Routing in Wireless Sensor Networks, Wireless Communication and Networking Conference, Volume: 3, Page(s): 1305 - 1310 Vol.3, 2004 [34] Raminder P. Mann, Kamesh R. Namuduri, Ravi Pendse, EnergyAware Routing Protocol for Ad HocWireless Sensor Networks, EURASIP Journal on Wireless Communications and Networking 2005:5, 635–644 [35] Vamsi Paruchuri, Arjan Durresi, Leonard Barolli, Energy Aware Routing Protocol for HeterogeneousWireless Sensor Networks, Proceedings of the 16th International Workshop on Database and Expert Systems Applications (DEXA’05), IEEE Computer Society, 2005 [36] Hung Le Xuan, Young-koo Lee and Sungyoung Lee, Two Energy-Efficient Routing Algorithms for Wireless Sensor Networks, 4th International Conference on Networking (ICN

[37]

[38]

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46]

[47]

[48]

[49]

[50]

[51]

[52]

[53]


2005) Proceeding, pp.698-705, ISBN: 3-540-25339-4, ISSN: 0302-9743 Wang Jin, Shu Lei, Jinsung Cho, Young-Koo Lee, Sungyoung Lee and Yonil Zhong, A Load-Balancing and Energy-Aware Clustering Algorithm in Wireless Ad-Hoc Networks, The 1st International Workshop on RFID and Ubiquitous Sensor Networks in conjunction with the 2005 International Conference on Embedded and Ubiquitous Computing (EUC'2005) Nagasaki, Japan, ISBN: 3-540-30803-2, ISSN: 0302-9743, LNCS 3823, pp. 1108-1117 Siva D. Muruganathan, Daniel C. F. Ma, Rolly I. Bhasin, & Abraham O. Fapojuwo, A Centralized Energy-Efficient Routing Protocol for Wireless Sensor Networks, IEEE Radio Communications, March 2005 Giljae Lee, Jonguk Kong, Minsun Lee and Okhwan Byeon, A Cluster based Energy Efficient Routing Protocol without Location Information for Sensor Networks, International Journal of Information Processing System, Vol 1, No 1, 2005 Chih-Wei Shiou, Frank Yeong-Sung Lin, Hsu-Chen Cheng & Yean-Fu Wen, Optimal Energy-Efficient Routing for Wireless Sensor Networks, Proceedings of the 19th International Conference on Advanced Information Networking and Applications (AINA’05), 2005 Uk-Pyo Han, Sang-Eon Park, Young-Jun Chung, An Efficient Energy Aware Routing Protocol for Wireless Sensor Networks, http://iec.cugb.edu.cn/WorldComp2006/ICW4674.pdf S. Mahlknecht, S. A. Madani and M. Roetzer, Energy Aware Distance Vector Routing Scheme for Data Centric Low Power Wireless Sensor Networks, IEEE International Conference on Industrial Informatics, 2006 Backhyun Kim, and Iksoo Kim, Energy Aware Routing Protocol in Wireless Sensor Networks, IJCSNS International Journal of Computer Science and Network Security, Vol.6 No.1, January 2006 Li Li, Dong Shu-Song, Wen Xiang-Ming, An Energy Efficient Clustering Routing Algorithm for Wireless Sensor Networks, The Journal of China Universities of Posts and Telecommunications, Volume 13, Issue 3, September 2006 Hossam Hassanein And Jing Luo, Reliable Energy Aware Routing in Wireless Sensor Networks, Proceedings of the Second IEEE Workshop on Dependability and Security in Sensor Networks and Systems (DSSNS’06), IEEE Computer Society, 2006 Henoc Soude and Jean Mehat, Energy Efficient Clustering Algorithm for Wireless Sensor Networks, Second International Conference on Wireless and Mobile Communications (ICWMC'06), IEEE 2006 Li Qing , Qingxin Zhu, Mingwen Wang, Design of a Distributed Energy-Efficient Clustering Algorithm for Heterogeneous Wireless Sensor Networks, Elsevier, Computer Communications 29 (2006) 2230–2237 Selvadurai Selvakennedy and Sukunesan Sinnappan, An EnergyEfficient Clustering Algorithm for Multihop Data Gathering in Wireless Sensor Networks, Journal of Computers, Academy Publisher, Vol. 1, No. 1, April 2006 Muhammad Mahbub Alam, Md. Mamun-Or-Rashid and Choong Seon Hong, QoS-Aware Routing for Sensor Networks Using Distance-Based Proportional Delay Differentiation (DPDD), International Conference on Next-Generation Wireless Systems ICNEWS 2006. [50] Ruay-Shiung Chang and Chia-Jou Kuo, An Energy Efficient Routing Mechanism for Wireless Sensor Networks, Proceedings of the 20th International Conference on Advanced Information Networking and Applications (AINA’06), IEEE 2006 Jin Wang, Jinsung Cho and Sungyoung Lee, A Hop-based Energy Aware Routing Algorithm for Wireless Sensor Networks, www.uclab.khu.ac.kr/resources/publication/C_161.pdf R Vidhyapriya , Dr P T Vanathi, Energy Efficient Adaptive Multipath Routing for Wireless Sensor Networks, IAENG International Journal of Computer Science, 34:1, 2007 Mao Ye, Chengfa Li, Guihai Chen and Jie Wu, EECS: An Energy Efficient Clustering Scheme in Wireless Sensor Networks, International Journal of Ad Hoc and Sensor Wireless Network, 2007.


941


[54] Marco Zimmerling, Waltenegus Dargie and Johnathan M. Reason, Energy-Efficient Routing in Linear Wireless Sensor Networks, IEEE International Conference on Mobile Adhoc and Sensor Systems, 2007 [55] R Vidhyapriya, Dr P T Vanathi, Energy Aware Routing for Wireless Sensor Networks, International Conference on Signal Processing, Communications and Networking, ICSCN 2007. [56] [56] R Vidhyapriya, Dr P T Vanathi, Conserving Energy in Wireless Sensor Networks, IEEE Transactions on Potentials, Vol 26, Sept/Oct 2007 [57] Zhiwei Zhao, Xinming Zhang, Peng Sun, Pengxi Liu, A Transmission Power Control MAC Protocol for Wireless Sensor Networks, Proceeding ICN '07 Proceedings of the Sixth International Conference on Networking, 2007, ISBN:0-76952805-8 [58] Jaejoon Cho, Sungho Kim, Heungwoo Nam, Sunshin, An Energy-Efficient Mechanism using CLMAC Protocol for Wireless Sensor Networks, Third International Conference on Networking and Services(ICNS'07), IEEE, 2007 [59] Ki Young Jang, Kyung Tae Kim, Hee Yong Youn, An Energy Efficient Routing Scheme for Wireless Sensor Networks, IEEE Computer Society 2007 [60] Gerhard P. Hancke, C. Jaco Leuschner, SEER: A Simple Energy Efficient Routing Protocol for Wireless Sensor Networks, SACJ, No. 39, 2007 [61] Ye Ming Lu and Vincent W.S. Wong, An Energy-Efficient Multipath Routing Protocol for Wireless Sensor Networks, International Journal of Communication Systems, Volume 20 Issue 7, July 2007 [62] Taewook Kang, Jangkyu Yun, Hoseung Lee, Icksoo Lee, Hyunsook Kim, Byunghwa Lee, Byeongjik Lee, Kijun Han, A Clustering Method for Energy Efficient Routing in Wireless Sensor Networks, Proceedings of the 6th WSEAS Int. Conf. on Electronics, Hardware, Wireless and Optical Communications, Corfu Island, Greece, February 16-19, 2007 [63] [63] Y. H. Wang, C. P. Hsu, Y. C. Lin, C. S. Kuo, and H. Y. Ho, A Routing Method by Reactive Energy Decision in Wireless Sensor Networks, 21st International Conference on Advanced Information Networking and Applications Workshops, AINAW’07, IEEE, 2007. [64] Siho Cha, Minho Jo, Jong-Eon Lee, Dae-Young Kim, Seokjoong Kang, Kuk-Hyun Cho and Nobok Lee, Hierarchical Node Clustering Approach for Energy Savings in WSNs, IEEE Computer Society, 2007 [65] Ming Yu, Kin K. Leung, and Aniket Malvankar, A Dynamic Clustering and Energy Efficient Routing Technique for Sensor Networks, IEEE Transactions on Wireless Communications, Vol. 6, No. 8, August 2007 [66] Monia Ghobadi, Kui Wu, Energy Efficient-Mobile Sink Geographic Grid Routing for Wireless Sensor Networks, Proceedings of the Thirteenth Annual International Conference on Mobile Computing and Networking, 2007 [67] Jisul Choe, Keecheon Kim, EADD: Energy Aware Directed Diffusion for Wireless Sensor Networks, International Symposium on Parallel and Distributed Processing with Applications, 2008 [68] Lan Tien Nguyen, Xavier Defago, Razvan Beuran, Yoichi Shinoda, An Energy Efficient Routing Scheme for Mobile Wireless Sensor Networks, IEEE International Symposium on Wireless Communication Systems 2008 [69] Tai-Jung Chang, Kuochen Wang1, Yi-Ling Hsieh, A Colortheory-based Energy Efficient Routing Algorithm for Wireless Sensor Networks, Journal of Computer Networks, Vol. 52, No. 3, pp. 531-541. (SCI, EI), 2008 [70] Jalil Jabari Lotf, Mehdi Nozad Bonab and Siavash Khorsandi, A Novel Cluster-based Routing Protocol with Extending Lifetime for Wireless Sensor Networks, 5th International Conference on Wireless and Optical Communications Networks, 2008 [71] Azeddine Bilami, Djallel Eddine Boubiche, A Hybrid Energy Aware Routing Algorithm for Wireless Sensor Networks, International Symposium on Computers and Communications, IEEE 2008 [72] Jianming Zhang,Yaping Lin,Cuihong Zhou and Jingcheng Ouyang, Optimal Model for Energy-Efficient Clustering in

[73]

[74]

[75]

[76]

[77]

[78]

[79]

[80]

[81]

[82]

[83]

[84]

[85]

[86]

[87]

[88]

[89]


Wireless Sensor Networks using Global Simulated Annealing Genetic Algorithm, International Symposium on Intelligent Information Technology Application Workshops, IEEE Computer Society, 2008 Rozeha A. Rashid, Wan Mohd Ariff Ehsan W. Embong, Azami Zaharim and Norsheila Fisal, Development of Energy Aware TDMA-Based MAC Protocol for Wireless Sensor Network System, European Journal of Scientific Research, Vol.30 No.4 (2009), pp.571-578 Navin Gautam, Won-Il Lee, and Jae-Young Pyun, Track-Sector Clustering for Energy Efficient Routing in Wireless Sensor Networks, IEEE Ninth International Conference on Computer and Information Technology, IEEE Computer Society, 2009 Haosong Gou, Younghwan Yoo and Hongqing Zeng, A Partition based LEACH Algorithm for Wireless Sensor Networks, IEEE Ninth International Conference on Computer and Information Technology, IEEE Computer Society, 2009 Haosong Gou and Younghwan Yoo, An Energy Balancing LEACH Algorithm for Wireless Sensor Networks, Seventh International Conference on Information Technology, IEEE Computer Society, 2010 Fuad Bajaber and Irfan Awan, Energy Aware Adaptive Clustering for Wireless Sensor Networks, International Conference on Network-Based Information Systems, IEEE Computer Society, 2009 Houda Zeghilet, Nadjib Badache and Moufida Maimour, Energy Efficient Cluster-based Routing in Wireless Sensor Networks, 14th IEEE Symposium on Computers and Communications, ISCC'09, IEEE 2009 Kavi K. Khedo, and R. K. Subramanian, MiSense Hierarchical Cluster-Based Routing Algorithm (MiCRA) for Wireless Sensor Networks, Proceedings of World Academy of Science, Engineering and Technology, Issue 52, April 2009 Noor M. Khan, Zubair Khalid and Ghufran Ahmed, GRAdient Cost Establishment (GRACE) for an Energy-Aware Routing in Wireless Sensor Networks, EURASIP Journal onWireless Communications and Networking, Article ID 275694. Hiren Kumar Deva Sarma, Avijit Kar and Rajib Mall, Energy Efficient Communication Protocol for a Mobile Wireless Sensor Network System, IJCSNS International Journal of Computer Science and Network Security, Vol.9 No.2, February 2009 M. Bani Yassein, A. Al-zoubi, Y. Khamayseh, W. Mardini, Improvement on LEACH Protocol of Wireless Sensor Network (VLEACH), International Journal of Digital Content Technology and its Applications, Volume 3, Number 2, June 2009 [83] Dilip Kumar A, Trilok C, Aseri B, R.B. Pate, EEHC: Energy Efficient Heterogeneous Clustered Scheme for Wireless Sensor Networks, Elsevier, Computer Communications 32 (2009) 662–667 Huang Lu, Jie Li and GuojunWang, A Novel Energy Efficient Routing Algorithm for Hierarchically Clustered Wireless Networks, Proceedings of the International Conference on Frontier of Computer Society and Technology, 2009 K S Shivaprakasha, Dr Muralidhar Kulkarni, Improved Network Survivability using Energy Aware DSR for Wireless Sensor Networks, Proceedings of IETE Conference on RF and Wireless, 8th and 9th Oct 2010, IETE Centre, Bengaluru Dilip Kumar, T. C. Aseri and R. B. Patel, EECHDA: Energy Efficient Clustering Hierarchy and Data Accumulation for Sensor Networks, BVICAM’S International Journal of Information Technology, 1-8, 2010 Basavaraj S.Mathapati, Dr.V.D.Mytri and Dr.Siddarama R. Patil, An Adaptive Energy Efficient Reliable Routing Protocol for Wireless Sensor Networks, ACEEE International Journal on Network Security, Vol 1, No. 1, Jan 2010 M.K.Jeya Kumar, Evaluation of Energy-Aware QoS Routing Protocol for Ad Hoc Wireless Sensor Networks, International Journal of Electrical, Computer, and Systems Engineering 4:3 2010 Saeed Ebadi , Ahmad Habibizad Navin and Mehdi Golsorkhtabar Amiri, Energy Efficient Cluster-based Routing Algorithm for Prolonging the Lifetime of Wireless Sensor


942


[90]

[91]

[92]

[93]

[94]

[95]

[96]

[97]

[98]

[99]

[100]

[101]

[102]

[103]

[104]

[105]

[106]

[107]

Networks, Journal of Global Research in Computer Science, Volume 1, No. 1, August 2010 Shio Kumar Singh, M P Singh, and D K Singh, Energy Efficient Homogenous Clustering Algorithm for Wireless Sensor Networks, International Journal of Wireless & Mobile Networks (IJWMN), Vol.2, No.3, August 2010 Kanojia Sindhuben Babulal, Rajiv Ranjan Tewari, E2XLRADR (Energy Efficient Cross Layer Routing Algorithm with Dynamic Retransmission for Wireless Sensor Networks, International Journal of Wireless & Mobile Networks ( IJWMN ), Vol.2, No.3, August 2010 Nan Ding, Guozhen Tan, Wei Zhang, Opportunistic Routing for Time-Variety and Load-Balance over Wireless Sensor Networks, Wireless Sensor Network, Vol 2, September 2010, PP 718-723 [93] Ehsan Ahvar, Mahmood Fathy, BEAR: A Balanced EnergyAware Routing Protocol for Wireless Sensor Networks, Wireless Sensor Network, Vol 2, October 2010PP: 793-800 [94] Seyed Mahdi Jamei, Proposing a New Energy Efficient Routing Protocol for Wireless Sensor Networks, IJCSNS International Journal of Computer Science and Network Security, Vol10 No.2, February 2010 Parul Saini and Ajay K Sharma, Energy Efficient Scheme for Clustering Protocol Prolonging the Lifetime of Heterogeneous Wireless Sensor Networks, International Journal of Computer Applications (0975 – 8887), Volume 6– No.2, September 2010 Ouadoudi Zytoune, Youssef Fakhri, Driss Aboutajdine, A Novel Energy Aware Clustering Technique for Routing in Wireless Sensor Networks, Wireless Sensor Network, Vol 2, March 2010, 233-238 [97] Surender Soni and Narottam Chand, Energy Efficient MultiLevel Clustering to Prolong The Lifetime of Wireless Sensor Networks, Journal of Computing, Volume 2, Issue 5, May 2010, ISSN 2151-9617 [98] Wang Jin, Imanishimwe Jean de Dieu, Asturias De Leon Diego Jose, Sungyoung Lee and Young-Koo Lee, Prolonging the Lifetime of Wireless Sensor Networks via Hotspot Analysis, IEEE Computer Society, 2010 Amir Hossein Mohajerzadeh, Mohammad Hossien Yaghmaee, Tree Based Energy and Congestion Aware Routing Protocol for Wireless Sensor Networks, Wireless Sensor Network, Vol 2, Feb 2010, 233-238 Yung-Fa Huang, Tung-Jung Chan, Tsair-Rong Chen, YoungLong Chen & Neng-Chung Wang, Performance of an Energy Efficient Routing Scheme for Cluster-based Wireless Sensor Networks, Journal of Networks, Vol. 5, No. 8, August 2010 Mrityunjay Singh, Niranjan Lal, Monika Sethi and Saroj Poonia, A Tree Based Routing Protocol for Mobile Sensor Networks (MSNs), International Journal on Computer Science and Engineering, Vol. 02, No.01S, 2010, 55-60 Young Han Lee, Kyoung Oh Lee, Hyun Jun Lee, Aries Kusdaryono, CBERP: Cluster Based Energy Efficient Routing Protocol for Wireless Sensor Network, Proceedings of the International Conference on Recent Advances in Networking, VLSI and Signal Processing, UK, February 2010 [103] K S Shivaprakasha, Dr Muralidhar Kulkarni, Praveen Kumar, Santhosh Kumar Singh, Energy Efficient Routing Protocol using Optimal Path for Wireless Sensor Networks, CiiT International Journal of Networking and Communication Engineering, Feb 2011, OCP Science. P. J. Hawrylak, L. Mats, J. T. Cain, A. K. Jones, S., Ultra-Low Power Computing System for Wireless Devices, International Review on Computers and Software (IRECOS), Vol 1, No 1, July 2006. L. Derdouri, C. Pham, M. Benmohammed, A Comparative Analysis of Reliable Multicast Protocols in Active Networking Environments, International Review on Computers and Software (IRECOS), Vol 3, No 3, May 2008. A. K. Dwivedi, Sunita Kushwaha, O. P. Vyas, Performance of Routing Protocols for Mobile Adhoc and Wireless Sensor Networks: A Comparative Study, International Journal of Recent Trends in Engineering, ACEEE, Vol 2, No. 4, November 2009 Hafedh Zayani, Rahma Ben Ayed, Wireless Sensor Networks Optimization: Cross-Layer (DSR-Z-MAC) and Synchronization

Technique (SMAC), International Review on Computers and Software (IRECOS), Vol 4, No 1, Jan 2009. [108] Kechar Bouabdellah, Sekhri Larbi, Rahmouni Mustapha Kamel, A Cross-Layer Design Approach to Achieve Energy Saving and Low Latency in MAC Layer for Delay Sensitive Wireless Sensor Network Applications, International Review on Computers and Software (IRECOS), Vol 4, No 3, May 2009. [109] Ali Hassoune M., Mekkakia Z. LAT-MAC: A Latency Energy Aware MAC Protocol for Wireless Sensor Networks, International Review on Computers and Software (IRECOS), Vol 5, No 3, May 2010.

Authors’ information K. S. Shivaprakasha received his BE (Electronics & Communication) from Bahubali College of Engineering, Visvesvaraya Technological University with IX rank to the University and MTech (Digital Electronics and Communication Systems) from Malnad College of Engineering, Visvesvaraya Technological University with I rank to the University in 2004 and 2007 respectively. Presently he is pursuing his PhD at National Institute of Technology Karnataka, Surathkal in the field of Wireless Sensor Networks. He has a teaching experience of 5 years. Presently he is a Senior Lecturer at Bahubali College of Engineering, Shravanabelagola, Karnataka, India. His areas of interest include Wireless Sensor Networks, Mobile Adhoc Networks, Information Coding Theory and Cryptography. He has more than 15 publications to his credit. Muralidhar Kulkarni received his B.E. (Electronics Engineering) degree from University Visvesvaraya College of Engineering, Bangalore University, Bangalore, M. Tech (Satellite Communication and Remote Sensing) from Indian Institute of Technology, Kharagpur (IIT KGP) and PhD from JMI Central University, New Delhi in the area of Optical Communication networks. He has 28 years of experience which includes 5 years in industry and 23 years of teaching experience. He has held the positions of Scientist in Instrumentation Division at the Central Power research Institute, Bangalore (1981-1982), Aeronautical Engineer in Avionics group of Design and Development team of Advanced Light Helicopter(ALH) project at Helicopter Design Bureau at Hindustan Aeronautics Limited(HAL), Bangalore (1984-1988), Lecturer (Electronics Engineering) at the Electrical Engineering Department of University Visvesvaraya College of Engineering, Bangalore (1988-1994) and Assistant Professor in Electronics and Communication Engineering (ECE) Department at the Delhi College of Engineering (DCE), Govt. of National Capital territory of Delhi, Delhi (1994-2008). He has served as Head, Department of Information Technology and Head, Computer Center at the Delhi College of Engineering (University of Delhi), Delhi. Currently, he is a Professor and Head in the Department of Electronics and Communication Engineering (ECE) Department, National Institute of Technology Karnataka (NITK), Surathkal, Karnataka, India. He is currently the Coordinator of the Centre of Excellence for Wireless Sensor Networks, Dept. of Electronics and Communication Engineeing, National Institute of Technology Karnataka. Dr. .Kulkarni’s teaching and research interests are in the areas of Digital Communications, Fuzzy Digital Image Processing, Optical Communication and Networks, and Wireless Sensor Networks. He has published several research papers in the above areas, in national and International journals of repute. For various contributions his Biography has been listed in the Marquis, Who's Who in Science & Engineering (2008). He has also authored/coauthored four very popular books in Microwave & Radar Engineering, Communication Systems, Digital Communications and Digital Signal Processing



943


Efficiency of MDB Routing Algorithm Over DB Routing Algorithm in Point-Point Networks S. Anuradha1, G. Raghu Ram2, T. Bhaskara Reddy3, J. Chakradhar4

Abstract – The goal of this paper is to investigate to control the congestion in the Communication Networks. The study is based on use of an ant algorithm operating upon a dynamic problem domain, that is, a domain that changes as a function over time. Specifically, discussed about the use of an ant-inspired graph-based general- purpose algorithm metaheuristic named Ant Colony Optimization as the basis of the method of implementation of a routing algorithm in a packet- switched point-to-point network (such as the Internet). Ant-inspired algorithms have the capability of finding short paths in graphs, and show an inherent adaptability that could be utilized to solve dynamic problems such as routing with node balancing in a network. Here a new general purpose heuristic algorithm named (MDB) Modified Depth-Breadth routing algorithm is defined and its performance was compared with (DB) Depth-Breadth routing algorithm and can be prove more efficient than DB routing algorithm. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: ACO, DB-Routing, MDB-Routing, Congestion Control

I.

In conjunction with a flow control, congestion and admission, routing determines the total network performance, in terms of quality and amount of offered services. The routing task is performed by routers, which update their routing tables by means of an algorithm specially designed for this purpose [5], [9]. The first routing algorithms addressed data in a network minimizing any costs function, like physical distance, link delay, etc [8], [10]-[12]. The goal of this paper is to focus on controlling the congestion in the Communication Networks. The study is based an ant algorithm which works on a dynamic problem domain. Specifically, discussed about the use of an antinspired graph-based general- purpose algorithm metaheuristic named Ant Colony Optimization as the basis of the method of implementation of a routing algorithm in a packet- switched point-to-point network (such as the Internet). Ant-inspired algorithms have the capability of finding short paths in graphs, and show an inherent adaptability that could be utilized to solve dynamic problems such as routing with node balancing in a network. [About DB routing and its implementation was given in the published papers in the same journal in Voume.4, n°3, May 2009].

Introduction

Modern communication networks are becoming increasingly diverse and heterogeneous. This is the consequence of the addition of an increasing array of devices and services both wired and wireless. The need for seamless interaction of numerous heterogeneous network components represents a formidable challenge, especially for networks that have traditionally used centralized methods of network control. This is true for both packet-switched and virtual circuit networks, and the Internet, which is becoming an ever more complex collection of a diversity of subnets [1],[6],[7]. The need to incorporate wireless and possibly ad-hoc networks into the existing wire-link infrastructure renders the requirement for efficient network routing even more demanding. Routing algorithms in modern networks must address numerous problems. Two of the usual performance metrics of a network are average time of traversal and average throughput. Current routing algorithms are not adequate to tackle the increasing complexity of such networks [2]-[3]. Centralized algorithms have scalability problems; static algorithms have trouble keeping up-to-date with network changes, and other distributed and dynamic algorithms have oscillations and stability problems [4], [10].Routing in a data network is the action of addressing data traffic between pair of nodes (source-destination), which is fundamental in a communication network control.

II.

Over-View of DB Routing

The By analyzing the two searching techniques DFS and BFS we came to a conclusion that each technique is


944


S. Anuradha, G. Raghu Ram, T. Bhaskar Reddy, J. Chakradhar

having its own advantages and disadvantages. By combining the key ideas of DFS and BFS we lead to a new technique which can take the DFS path or BFS path to converge to the better solutions than DFS and BFS and we name this new technique as DB (since it routes Depth wise, Breadth wise in the tree of nodes) and leads to better paths as solution.

output flows is less than or equal to the sum of its input flows: 1.1 Assign starting flows to network edges 1.2 Initialize the list of nodes expected to be processed N=P 2 do 3 Take the first node P from the list N 4.1 Recalculate output flows of node P 4.2 Add all nodes that have been affected by this action to list N 5 while(S 6= {})

III. MDB Routing Algorithm The Original DB routing algorithm focuses only on the searching method used in finding out the shorter paths in the network but not on node balancing concept. Based on the original DB routing algorithm few modifications are suggested to be added to enhance its performance and named this as modified Depth-Breadth routing algorithm.

III.3. Experimental Setup The simulated network is the domain in which the implemented routing algorithm with combined searching technique – presented in this paper – works. By modeling a network and letting the algorithm work with representations of objects in a physical network, we were forced to think in terms of what is possible in a “real” network. The algorithm that have implemented requires a few functionalities in the network protocol that are not commonly in use. Routing tables: The routing tables implemented for our algorithm are organized in this way and contain entries that hold a trail-value that is used in the stochastic selection process. Ants are packets: All the packets, which are used to route in MDB, are ants. As every ant maintains a memory that is used as a taboo-list, and for retracing the path to update trail-values, this memory has to be stored in the packet – either as payload or as a designated header-field. Instant router update: A delayed update of trailvalues was used, so information regarding the quality of an ant’s path should somehow be propagated back through the network to the routers on the used path. In the implementation, this propagation is not effectuated, only simulated. Instead, all routers are instantly informed of the quality of the path by the metaheuristic.

III.1. Congestion Control (Control of the Number of Ants in the Network) For every algorithm, the network load generated by routing packets is stored as the ratio between the bandwidth occupied by all the routing packets and the total available network bandwidth. The routing overhead is the main function of the topological properties of the network and of the generation rate of the routing information packet. Ant Net produces a routing overhead depending on the ants’ generation rate and the length of the path along they travel. As the followed path of routing ant grows (either because of topology or bad routing) the routing overhead grows. The DB routing and original Ant Net does not take into account the generated routing overhead and its effect on overall network performance. III.2. Introducing Load Balancing Technique to Overcome the Above Mentioned Problem Original DB routing algorithm addresses the routing problems but not load balancing [11]. Load balancing is heavily relied on routing; Ant Net routing philosophy can lead to network congestion, high delay and may create deadlock. For example a node that lies on several routes will have a large number of packets for different destinations in its interface queue; all these packets will experience high queuing delays resulting in a high overall end to end delay. Load balancing technique is needed to remove such bottlenecks. The optimal solution found in the first phase is then finalized by a deterministic procedure that adjusts flows in order to achieve the precise balance of the inputoutput flow at each node. A outline of this procedure is shown as follows: Input: Original flows stored Output: Modified flows that satisfy the condition in which each inner node of the network, the sum of its

IV.

Event Driven Network Simulation

To facilitate an empirical study of the behavior of ant inspired routing algorithms in the proposed network model, it was necessary to simulate the movement of packets over time in the model. To this end we have designed and implemented an event driven simulation scheme. The simulation is an adaptation of the event driven simulation scheme proposed in Weiss 1998 [13], where a priority-queue ensures that events are executed in order. IV.1. The Concept of Event Location An event-driven simulation only needs to contain a number of events with a specified time, and an eventholder, which handles the scheduling of these events. We



945


have a system, where the defined events are spawned by other events. This is simply because a packet travels in a router-wire-router-wire-… cycle. The complexity in the system lies in knowing when a packet is ready for a switching operation or a sending operation (at a router or wire respectively), because of the buffering in the inputand output queues.

connected in a grid with wires (with a delay of 10) between adjacent routers. This topology is used to test DB ability to route in a ‘larger’ (but still simple) network.

VI.

VI.1. Performance of Modified Depth-Breadth Routing Algorithm on 4X4 Topology with Different Network Loads

IV.2. Simulation Events The events that can occur in this event-driven simulation to be defined. The goal here is to simulate routing in our network model, a task requiring several steps that has already been discussed. Briefly outline the events that have defined, their duration and purpose. The events and their properties should follow naturally from the description of routing in the network model and the discussion of packet loss. The packet creation is initiated by an input queue event. This should be interpreted as if an external entity is responsible for creating the packet and has sent it to the initial router in our model. This mean that all new packets must wait in the input queues, and they risk being lost if the queue capacity is exceeded.

V.

In this section different test runs were conducted on 4X4 topology with 16 nodes and the results were displayed in the following Figs. 2 to 7 and Tables I & II and graphs in the Figs. 8 & 9. The study effects on increase in packet load which increases the average time for the packet transmission in milliseconds on the topology this is obvious because of the increase in load in the network. Three different experiments were conducted on both MDB and DB routing algorithms. Starting with the initial network load as 3500 and later with 5000 and 10000 packets. The comparative results for the experiments are placed in Fig. 8 with No. of packets Vs Average time and in the Fig. 9 with No of packets Vs No of dead packets.

Network-Topology

The grid network topologies were chosen as an example of a ‘large’ network in which it is easy to investigate and describe traffic, since it has a regular structure. need a regular topology because of the ‘reduced’ complexity and simply because it is easier to get a general view of the paths in the network. The Grid networks are a special case of Communication networks. Since they are easy to setup and maintain, and have good scalability, GCNs are potentially a popular access method for hospitals, hotels, and conference centers. This studies routing algorithm for grid networks, using MDB routing, which addresses load-balancing on grid topologies. The experiments presented in this proposed work have been made on special network topology 4X4 grid.

Experimental Analysis

1

2

4

5

6

8

13

14

16

Fig. 2. The paths generated from experiment 1 by DB routing algorithm

Fig. 1. The 4x4 Grid network

V.1.

4X4 grid (16 nodes)

It had furthermore performed experiments on a 4X4 grid network, which of course consists of 16 routers

Fig. 3. The paths generated from experiment 1.1 by MDB



946


Fig. 4. The paths generated from experiment 2.1 by DB routing algorithm

Fig. 7. The paths generated from experiment 3.2 by DB routing algorithm TABLE II RESULTS FROM EXPERIMENTS 1.2, 2.2, 3.2 BY DB ROUTING ALGORITHM Packet Load No. Of Dead packets

Average time for the Packet

3500

694

316.13

5000

3429

239.79

10000

4998

219.89

Fig. 5. The paths generated from experiment 2.2 by MDB

Fig. 8. Comparing results of Number of packets Vs Average time for packet in 4X4 grid using MDB and DB algorithms

Fig. 6. The paths generated from experiment 3.1 by MDB routing algorithm TABLE I RESULTS FROM EXPERIMENTS 1.1, 2.1, 3.1 BY MDB ROUTING ALGORITHM Packet Load No. Of Dead packets

Average time for the Packet

3500

494

306.53

5000

2492

219.77

10000

4791

199.84

Fig. 9. Comparison of MDB and bench mark algorithms for No. of packets Vs No. of dead packets in 4X4 grid



947


VII.

Conclusion [5]

From the above test runs executed on 4X4 grid using DB-Routing and MDB routing algorithms respectively, with the generated paths one can view all the heuristics of ACO (ant colony optimization) which were assumed for the practical implementation in the MDB routing algorithm were true and successful in executing can be visualized in the Fig. 2 through 7. By observing the statistics from the Tables I and II as the no of packets were increasing, the scope for the balancing was more and the performance of MDB was more improved with 5000 than 3500 and 10000 than 5000 this can be concluded from the average time of the packet traversal and it performing better than DB routing as there is an improvement in No of dead packets and average time. For 3500 as packet load with MDB the average time for the packet traversal from source to destination was 306.53 milliseconds and the no of dead packets was 494 where as it was 316.13 average time and 694 is the dead packet count. There is an improvement of 9.6 milliseconds of average time and 200 dead packets over DB routing algorithm. With 5000 as packet load with MDB the average time for the packet traversal from source to destination was 219.77 milliseconds and the no of dead packets was 2492 where as it was 239.79 average time and 3429 is the dead packet count. There is an improvement of 20.02 milliseconds of average time and 937 dead packets over DB routing algorithm. With 10000 as packet load with MDB the average time for the packet traversal from source to destination was 199.84 milliseconds and the no of dead packets was 4791 where as it was 219.89 average time and 4998 is the dead packet count. There is an improvement of 20.50 milliseconds of average time and 207 dead packets over DB routing algorithm. From the graphs plotted for both the algorithms one for Average time for the packets and other for Dead packet count which were depicted in the Figs. 8 and 9. From the statics and the graphs we can conclude that MDB-routing is efficient than DB routing algorithm with minimum dead packet count and average time for the packet traversal.

[6]

[7] [8]

[9]

[10]

[11]

[12]

[13]

Authors’ information 1 Associate Professor of M.C.A and Additional controller of Exams, G. Pulla Reddy Engineering College, Kurnool, A.P, India. 2 Head Of M.C.A Department, G. Pulla Reddy Engineering College, Kurnool, A.P, India. 3 Associate Professor in the department of Computer science and Technology in Sri Krishna Devaraya University, Anathapur, A.P, India. 4 Associate Professor and R&D Incharge, in Sri Venkatesa Perumal College of Engineering and Technology, Puttur, AP, India. Andhra Pradesh, India.

Dr. S. Anuradha obtained her B.Sc and MCA degrees from Osmania University, Hyderabad in the year 1994 and 1997 respectively. Awarded PhD in March 2011at Sri Krishna Devaraya University, Anantapur, India. She is presently working as Associate Professor in the Department of Master of Computer Applications and Additional controller of Exams at G. Pulla Reddy Engineering College, Kurnool, Andhra Pradesh, India. She presented more than 15 research papers in various national/ International journals and conferences. Her research areas include Computer Networks and Routing Algorithms. E-mail: [email protected]

References [1]

[2]

[3]

[4]

A Stateless Active Queue Management Scheme for approximating Fair Bandwidth Allocation. IEEE INFOCOM, Mar 2009 P. E. McKenney. “Stochastic Fairness Queueing”. Proc. IEEE INFOCOM, 2:733– 740, June 1990. De Marco, F. Postiglione, M. Longo, ―Run-time adjusted congestion control for multimedia: experimental results, Journal of Interconnection Networks (JOIN), vol. 5, no. 3, pp. 249-266, 2008. Park and W. Willinger, Self-Similar Network Traffic and Performance valuation, John Wiley and Sons, 2000. C. Blum. Ant colony optimization for the edge-weighted kcardinality tree problem. In GECCO 2002: Proceedings of the Genetic and Evolutionary Computation Conference, pages 27–34, 2002. S.Anuradha, G.Raghuram, K.E.Sreenivasa murthy, V.Raghunath Reddy, B.Satyanarayana “Evolution Of New Searching Technique In Unicast Routing Algorithm Of Data Networks”, International Review on Computers and Software (IRECOS) Voume.4, n°3, May 2009,pages(321-330). S.Anuradha, G.Raghuram, K.E.Sreenivasa murthy, V.RaghunathReddy, B.Satya narayana “Genesis of DB Routing Algorithm in Unicasting Networks” ICGST Inter National Journal on computer Networks and Internet Research(CNIR) Volume 9, Issue I, July 2009, Pages (47.. 55). S.Anuradha,G.Raghuram,K.E.Sreenivasamurthy,V.Raghunath Reddy “Event Driven Network Simulation in Combined Searching Technique Used to Speedup Packet Transmission in Computer Networks” IEEE International Conference on Methods and Models in Computer Science (ICM2CS-2009) on Dec 14-15th 2009 at School of Computer and Systems Sciences Jawaharlal Nehru University, New Delhi-110067. G.RaghuRam,S.Anuradha,Dr.V.Raghunatha Reddy,Dr.T.Bhaskar Reddy,’Setting Parameters for Basic Properties of a Node Balancing Protocol in Point-Point Network Communications’ IJCSES International Journal of Computer Sciences and Engineering Systems, Vol.5, No.1, January 2011. G. M. Weiss and H. Hirsh, Learning to Predict Rare Events in Event Sequences, In Proceedings of the Fourth International Conference on Knowledge Discovery and Data Mining, AAAI Press, Menlo Park, CA, 359-363, 1998.

S.Anuradha, G.RaghuRam, K.E.Sreenivasamurthy, V.Raghunath Reddy, D.R.Sreenivas “DB Routing algorithm- A new approach in Unicasting Networks To Speedup data Transmission” Academy Publisher’s International Joint Journal of Recent trends in Engineering ( CEE-2009, November 2009 issue). W. Chung, ―Congestion control for streaming media, Ph. D. dissertation, Polytechnic Inst., Worcester, 2005. Vol. 9 Issue 5 (Ver 2.0), January 2010 Global Journal of Computer Science and Technology. Sally Floyd and Kevin Fall, ―Promoting the use of end-to-end congestion control in the Internet, IEEE/ACM Transactions on Networking, vol. 7, no. 4, pp. 458–472, Aug. 1999. Rong Pan, Balaji Prabhakar, and Konstantinos Psounis. CHOKe,



948


J. Chakradhar Awarded Diploma in Computer Engineering - D.CM.E from Loyola Polytechnic in the year 2003, graduated in Information Technology - B.Tech from Jawaharlal Nehru Technological University Hyderabad in the year 2006, Awarded Masters of Computer Science and Engineering - M.Tech in the year 2010 from Jawaharlal Nehru Technological University

G. Raghu Ram graduated from Sri Krishna Devaraya University in the year 1994, MCA from Osmania University in the year. He is presently Associate Professor in the Department of Master of Computer Applications at G. Pulla Reddy Engineering College, Kurnool, and Andhra Pradesh, India. He published more than 13 papers in various national and international Journals and conferences. His research areas include Artificial Intelligence and Computer Networks. E-mail: [email protected]

Anantapur. He is presently Associate Professor and Inchrge for Research & Development at Sri Venkatesa Perumal College of Engineering and Technology, Puttur, AP, India. He published 5 papers in various National /International JournalsHis area of interest includes Computer Networks, Information Security, Network Security and Operating Systems. E-mail: [email protected]

Dr. T. Bhaskara Reddy graduated in science from Sri Venkateshwara University in the year 1984 Tirupathi, AP, India. Post graduated in M.Sc and M.Tech in 1997 and 2008 respectively. Obtained Ph.D in 2006 from Sri Krishna Devaraya University Ananthapur, AP, India. He is Presently Associate professor in the department of Computer science and Technology, Dy.Director distance education at Sri Krishna Devaraya University, Ananthapur, AP, India, and Published 46 papers in various National and International Journals and conferences. His research area includes Computer Networks and Data ware housing. E-mail: [email protected]



949


Real Time “2CRT” Architecture and Platform for the Experimentation of Telecommunication Terminals According to the Manhattan Mobility M. El Bakkali1, H. Medromi2

Abstract – In the field of technology the verification of the robustness of a system is very important. This system verification compares the actual with the origin bearing in mind that any anomaly could lead to a catastrophe. The control could be in different forms: either measure, or test the first control is done to measure the physical parameters in order to confirm system reliability as it is the case of « 2CRM » (Connection, Control, Recognition and Measure) [1]. The second control enables a remote experimentation to certify credibility of the terminal. Our article presents a new architecture of control « 2CRT » (Connection, Control, Recognition and Test). Its newness is based on the distribution of actions as it is based on multi-agent systems. The control is done remotely via the web using Real Time constraints. It meets to the functionalities of distribution, cooperation and adaptation to the changes in user behavior, environments and materials. Notably, the approach includes the test control mode of telecommunication terminals. The solution is based upon the Manhattan mobility with the possibility of integrating other topologies. This experimentation is done via the web by matrix configurations using FIFO (First In First Out) data structures. To further embody our architecture, we perfected it in the OpenAirInterface project of the Eurecom laboratory in France. Our control platform was developed using the unified process « RUP » (Rational Unified Process) [2] with the use of the UML modelization to well illustrate the system. It is implemented by an open source distribution (Java and J2ee running on Eclipse). Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Measure Control, Test Control, Distributed Systems, Real Time, 2CRM, 2CRT, MultiAgent Systems, Physical Parameters, Labyrinth, Telecommunication Terminals, Manhattan Mobility, Matrix Configurations, FIFO, Openairinterface, RUP, UML, J2ee

I.

PS: The equipment is of course essential elements to manage the transmission of signals from a transmitter to a receiver [3] [4].

Introduction

The test and experimentation platform regroups a set resources and technical means to aid businesses and researchers to prototype and design their services and innovative content upon a target technology. It is also a center for watching over the strategies of companies involved in the value chain of networks, and an observatory of emerging uses. Many fields (as telecommunication, transport, etc.) are still innovative especially with the technological and popular growth. In order to achieve the same model of growth, this technology ought to be in the control phase. Those controls must allow the verification of the inter-working between an application in a terminal and an application in a server. This verification implies the monitoring of the interactions caused by a request sent by the terminal. Hence, it is a test of the services interoperability in the overall communication architecture.

II.

The Existing Control Systems

The existing test platforms dedicated to test interworking are the following: the TIPHON project [5] [ 6] of the ETSI, which deals with large scale inter-working tests; the test generating platform designed by Korea Telecom [7] and the Information and Communication university of Korea, and the test platform of the Tsinghua university in Beijing China. France has also developed conform test generating tools: for example the Test Composer tool of the company Verilog-Telelogic and the TGV tool [8] of l' IRISA. The above mentioned test platforms are partial: in the case of TIPHON, it is a project of large scale experimentation platform which ignores validation. This platform has not been implemented yet. And in the case


950


M. El Bakkali, H. Medromi

of Korean and Chinese colleagues, the functionalities offered by their platforms are limited, either because they test only control aspects without considering the data or either because some steps are not fully developed as the test execution for example. The test generation tools don't have the functionalities of a platform [9].

second version "OpenAirInterface 2" which is under development. In the case of the first version of OpenAirInterface, the platform is more based upon Wireless Mesh Networks (WMN) [13] (Fig. 2).

III. Description of the Proposed Architecture The 2CRT (Connection, Control, Recognition and Test) architecture allows the connection to any server that is already stored in the knowledge base in order to operate it into performing tasks. This architecture achieves a series of actions which starts with a well chosen initial situation and leads to the desired goal. Hence, one can say that the result of the planning is a plan [10]. This approach is intended primarily to Real Time computer systems by taking into account the time constraints with respect to the accuracy of the result [11] (Fig. 1). PS: A 2CRT task could be executed via Internet during the application of test manipulation

Fig. 2. Mesh network [12]

The OpenAirInterface2 platform is, in the other hand, more based upon cellular networks (communications networks specifically designed for mobile devices.) [14] (Fig. 3).

Fig. 3. Cellular network [12]

Consequently, our architecture will enable the researchers and companies to better visualize the OpenAirInterface platform without writing into the configuration files. Especially in large scale networks in which it is very time consuming (Fig. 4). Fig. 1. "2CRT" Control architecture

IV.

Applying the Architecture to the OpenAirInterface Platform

The OpenAirInterface [12] is an open source hardware/software development platform, It is also an innovation forum in the field of digital communication. The OpenAirInterface was created by the Mobile Communications department of the EURECOM and is conducted primarily through collaborative projects (French projects ANR, European IST ...). The OpenAirInterface is still innovative since it is still under rigorous research work. This research leads to the development of a full version "OpenAirInterface 1” and a

Fig. 4. Schematic diagram of test control



951


Since the second version is still yet under development and depending to the integration needs of different mobilities such as Aléatoire, Manhattan, and Hexagonal …etc. We based our work upon the Manhattan model as a first step towards mobility integration. [15] This architecture (Fig. 5) is divided into two parts for both versions of OpenAirInterface: The first part is software installed in the OpenAirInterface server platform which allows local researchers to configure the platform and create a .top file with the desired settings. The second part is web based and allows external researchers (or companies) direct access to remotely configure and test the platform while putting the first part in Real Time. PS: In the first version the configuration could be done in one or many CH (Cluster Head) either physically or via visualization.

Example, when a request for an external user using a request agent, the latter will trigger an queuing agent and according to existing requests it will be in action when his turn will came as it is the case of FIFO "First In First Out." And so on for every action.

V.

UML Modeling

We identify a preliminary description of the uses case and its actors [17]. The planning of the objectives: • Intention: Plan via Internet a user request to perform a configuration of a set of telecommunication terminals; • Intention: the receipt of requests in Real-Time of for each terminal in use. • Actions: establish a request to a local (LAN) or remote (web) user • Actions: the server receives the request • Actions: the server verifies the availability of the terminals. • Actions: the server allocates the terminals • Actions: the server establishes a network between the terminals according to the requested configuration. • Actions: The server retrieves the configuration settings. • Actions: The server displays the settings to the user. In such a way, the actors and actions are graphically represented in a use case diagram (see Fig. 7).

Fig. 5. Detailed "2CRT" Control architecture

The integration of agents is important to make our architecture penetrating. [16] Consequently each task done by an agent goes into action. In return the objectives will be confidentiality and intelligence (Fig. 6). Fig. 7. Uses Case diagram of the "2CRT"

By exploiting the textual description of the use case already established, and analyzing all the scenarios, we can show the class diagram as shown in Fig. 8.

Fig. 6. Detailed "2CRT" Agent based Control architecture Fig. 8. Class diagram of the "2CRT"



952


There are associations from the sever class which have for objective the realization and execution of the configuration task requested by the user.

VI.

Software Architecture

This is software architecture for allocation request and testing. Technical explanation The user sends a request; a verification of the inserted data is done automatically. If the verification does not trigger any errors, the request is added to the "Liste_Requests.xml” file (this file is linked in Real Time with the application Sim-OpenAirInterface whatever the version 1 or 2) and the "Requests_n.xml” file which contains the requested data is automatically created. Only the application will be aware of a request, it will process it by configuring the TOP files in the OpenAirInterface-1 version and by relying on the Manhattan topology in the case of the second version OpenAirInterface-2. Once this processing is performed within the OpenAirInterface-Sim application, XML files containing the positions of the base stations and configuration files are created. Those files are then executed in OpenAirInterface (Fig. 9).

Fig. 10. Insertion request

"Save File XML" allows saving the positions of the eNB and the UE and the current proposition in the form of an XML file (Fig. 11), with the capitations: " Position_ENB.xml" and" Position_UE.xml ".

Fig. 11. The request folder

Fig. 12 allows viewing the random drawing, with the possibility to save it in the JPG format.

Fig. 9. "2CRT" Software architecture

VII.

Implementation

The insertion of data X and Y allows to draw a grid based on the number of X's and Y's, knowing that at any point (X, Y), there is a base station "eNB / CH" (Red Spot). Given that, L is the distance between a base station and another. Insertion of the number "UE" (User Equipment). After clicking on "Random Proposition", the positions of components eNB and UE appear, knowing that the UE are chosen randomly. The “Generate Randomly” button displays the same data as “Random Proposition" but with animation, expressing the random positions of UE (Fig. 10).

Fig. 12. Example of Manhattan topology

Fig. 13 shows that it is possible to modify the distance between base stations.

Fig. 13. Distance between base stations



953


"Real Time" allows to generate the requests of the web portal (Fig. 14).

Fig. 17. Results of the Sim-OpenAirInterface 2

VIII. Conclusion

Fig. 14. Real Time simulation

We applied in this article the "2CRT" architecture (cognizance , connection, control and test) in a relatively adjusted manner in the "OpenAirInterface". It aims to offer dedicated services of local or remote experiments which allow the evaluation of new mobile broadband services. These experiments are made available to research centers and service providers to simulate and configure this platform around the world, analyzing the connections and configuration via a web portal (developed in J2EE) and a configuration application (in Java). The proposed application has been tested and approved in the research laboratory of Eurecom in France. The advantage of our platform is that it can be integrated in different areas. Of course, this work is far from being completed. The second version of the OpenAirInterface platform is under development, paving the way for many new applications and perspectives. With regard to current architectures, the "2CRT" architecture offer a coordinated, coherent and proactive environment in order to control the reliability of the hardware and rapidly react in case of instability problems. The definition of the rules in the platform is made at the interface agents level, which facilitates the platform and improves its clarity. The model is based on software capabilities, allowing dynamic exchange between mutually cautious agents. Not to mention that this model can be personalized and applied in other areas than telecommunication, which is not the case of the current architectures.

The web page allows the insertion of the data (X,Y, the number of the EU)after a click on " Generate Randomly" button the processing and retrieval of results are done within the "Real Time " of the application, and not from the web page (Fig. 15).

Fig. 15. The Sim-OpenAirInterface 2 web portal

Data processing takes 4 seconds locally, after that the remaining time is to display the results on the net (Fig. 16).

References [1]

Fig. 16. Sim-OpenAirInterface 2 processing

[2]

Finally, the results of the eNB, UE positions, and the image of the Manhattan topology, are displayed (Fig. 17).

[3] [4] [5]


M. El Bakkali, H. Medromi / A Multi-Agent Systems based on a Real-time distributed “2CRM" Robot Platform for the measurement of telecommunications Lines and terminals, International Review on Computers and Software (IRECOS), November 2011. Philippe Kruchten / Introduction to Rational Unified Process, Eyrolles - January 2000. Guy Pujolle /Networks – 2008. Claude Servin / Networks and Telecoms – 2003. TIPHON / www.etsi.org/TIPHON - 2007.


954


[6] [7] [8]

[9]

[10]

[11] [12] [13] [14] [15]

[16] [17]

Vilho Räisänen / Implementing service quality in IP networks 2003. Marcien Mackaya / Test of the protocols and services of mobility, doctoral thesis at bordeaux I university - March 2005. Claude Jard and Thierry Jéron / TGV: theory, principles and algorithms - Integrated Design and Process Technology, Printed in the United States of America, June, 2002. Platonis is a multi-protocol and multi service experimentation and validation platform of the Réseau National de Recherche en Télécommunications « RNRT » et it's partners: INT, LabriBordeaux university, Limos- Univ. Blaise Pascal, Clermont Ferrand, France Telecom R&D et Kaptech. http://www-lor.intevry.fr/platonis/resume.html H. Medromi, F. Qrichi Aniba, A. Sayouti. “Real Time Distributed Architecture based on Multi-Agent System”. Second edition of the IT and decisional Math forum (JIMD'2008), ENSIAS, Rabat, Morocco, 3-5 July, 2008. Francis Cottet, Emmanuel Grolleau / Real Time ControlCommand systems, Conception and implementation – 2005. Software/Hardware open source development platform www.OpenAirInterface.com George Aggélou / Next Wireless Mesh Networking – 2009. Wale soyinka / Wireless Network Administration, A Beginner’s Guide – 2010. Jérôme Härri, Marco Fiore, Fethi Filali, Christian Bonnet / Vehicular Mobility Simulation with VanetMobiSim, The Society for Modeling and Simulation International (http://sim.sagepub.com/) - Janvier 2009. Barbara Dunin-Keplicz, Rineke Verbrugge / Teamwork in multiagent systems, A Formal Approach - 2010. Pascal Roques, Franck Vallée / UML 2 in action from needs analysis to the design, fourth edition– 2007.

Authors’ information Mohammed El Bakkali got his Master’s Degree in Computer Multimedia Telecommunications from Mohammed V Agdal University, Rabat, Morocco, in 2008. He obtained a degree in Electronics from Moulay Ismail University, Meknes, Morocco, in 2006. Since 2008, he has been preparing his PhD within the system architecture team of the ENSEM, Casablanca, Morocco. His actual main research interests concern the Design and Construction of a Measurement Robot Platform of Physical Parameters of Lines and Telecommunication Terminal, Real-Time Based on Multi-agents Systems.agents Systems. Hicham Medromi obtained the PhD in engineering science from the Sophia Antipolis University in 1996, Nice, France. He is responsible of the system architecture team of the ENSEM Hassan II University, Casablanca, Morocco. His actual main research interest concern Control Architecture of Mobile Systems Based on Multi Agents Systems. Since 2003 he is a full professor for automatic productic and computer sciences at the ENSEM, Hassan II University, Casablanca, Morocco.



955


A Bandwidth Request Competition Mechanism in WiMAX System Jianbo Ye, Qingbo Zhang

Abstract – Bandwidth competition resolution mechanism is a key technology of WiMAX. According to WiMAX agreement, Bandwidth competition resolution mechanism adopts the binary exponential return algorithm generally. But this algorithm will reduce the loss rate of bandwidth request and increase the number of SS and the loss rate bandwidth request as the initial return window increased. This paper proposes a bandwidth competition system based on a request to wait in WiMAX system. The proposed algorithm limits the number of bandwidth request sent in one frame, that is to say each user does not send new bandwidth request immediately after a request is successfully sent, but rather to wait for several frames before sending. The simulation results show that the proposed algorithm not only the method is simple, but also it can effectively reduce the loss rate and delay of bandwidth request. So, the proposed algorithm has practical engineering meaning. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: WiMAX, Bandwidth Competition, a Request to Wait, Loss Rate, Delay

I.

bandwidth request process to SS, which is called polling. These distributions may be to a single SS, may also be given to a group of SS. The polling of WiMAX is divided into three categories which are unicast, multicast / broadcast and PM bit. The unicast polling and PM are allocates sufficient bandwidth for the SS and poll the SS separately, and this does not appear the case which caused by the competition of the bandwidth request. While in the multicast / broadcast polling, BS in each frame is assigned a bandwidth for all SS. The SS which is participated in the multicast / broadcast polling send their respective bandwidth request during this period of bandwidth. It causes the conflict of competition and BS can’t correctly receive the bandwidth request message of SS. The time interval which occupied by the bandwidth is called the contention window. Contention window consists of many transmission opportunities. Each SS can be obtain the frame structure information by uplink mapping message, and sends a bandwidth request independently.

Introduction

WiMAX is the abbreviation for World Interoperability for Microwave Access which is a kind of wireless metropolitan area network access technology base on the IEEE 802.16 .It is proposed a new air interface standard which can realize broadband access with the last mile. Bandwidth competition resolution mechanism is a key technology of WiMAX. According to WiMAX agreement [1],[2] the SS ( Service Station ) using the binary exponential return algorithm to reduce competition conflict, and its initial return window and maximum return window are controlled by BS(Base Station),and notify the SS through the uplink. But the algorithm will reduce the loss rate of bandwidth request and increase the number of SS and the loss rate bandwidth request as the initial return window increased. Based on this, a lot of the improved algorithm is presented. The global optimal return window mechanism which sets a global optimal initial return window in each competition is proposed by reference [3]. The value of initial return window is equal to the number of SS which is participated in the competition. Let initial return window is equal to the maximum return window, thus the utilization rate of transmission chance can achieve to the maximum. This paper presents a new method based on a request for bandwidth competition mechanism.

II.

III. The Request Waiting Mechanism From above, it increases the delay of bandwidth request caused by the special regulations of SS behavior in the global optimal Rollback window Mechanism, while the delay of the improvement is not obvious. In order to reduce the collision times and the loss rate and delay of bandwidth request, it must limit the number of bandwidth request sent in one frame, that is to say each user does not send new bandwidth request immediately after a request is successfully sent, but rather to wait for several frames before sending.

Bandwidth Request Competitive Mechanism of WiMAX

The competition mechanism diagram is shown as Fig. 1. WiMAX agreement shows that BS distributes the Manuscript received and revised October 2011, accepted November 2011

956


Jianbo Ye, Qingbo Zhang

Reduce the time interval E (T ) can effectively reduce the transmission delay of SS data packet. Thus, the minimum value E (T ) can be obtained as: 1⎞ ⎛ m = n ln ⎜ 1 − ⎟ − 1 C ⎝ ⎠

(6)

Because n can’t be negative, we can obtain as above: ⎧ ⎪ 1⎞ ⎛ ⎪n ln ⎜ 1 − C ⎟ − 1 ⎝ ⎠ ⎪⎪ m=⎨ ⎪ ⎪0 ⎪ ⎪⎩

Fig. 1. The diagram of competition mechanism

Suppose ps is the bandwidth request probability 1 competition to ps send success. Then the bandwidth request probability in one frame is as follows:

which a SS sends successfully. It takes

1 ps 1 pr = = 1 + t 1 + ps t ps

(1)

n 1 + ps t

(2)

where， n is the number of SS which participates in the competition. Thus, the number of bandwidth request SS which a frame sent reduces to nr ,then ps can be gotten as above: 1⎞ ⎛ ps = ⎜ 1 − ⎟ ⎝ C⎠

nr −1

n

1 ⎞1+ p t ⎛ = ⎜1 − ⎟ s ⎝ C⎠

−1

(3)

n

IV.

(4)

Thus we can get the time interval which SS sents a bandwidth request successfully two times as follows: n

1 en 1

en

(7)

−1

(8)

Simulation Analysis

The binary exponential return mechanism, the global optimal return window and the request waiting competition mechanism proposed by this paper are compared using by Matlab. The binary exponential return mechanism is called the" basic" mechanism, the global optimal return window mechanism is referred to as "the optimal return" mechanism, the request waiting competition mechanism is called the "request waiting" mechanism. Parameters of simulation are set as follows. Simulation time is 1000 frames. For basic mechanism, the initial value of return window is 32 and the maximum return window is 1024. For the optimal return mechanism and the request waiting mechanism, the initial value of

−1

1 ⎛ 1 ⎞1+ m E (T ) = + t = ( m + 1) ⎜ 1 − ⎟ ps ⎝ C⎠

C>

−1

ˆ ) wopt = max ( n,c

where, C is the size of competition window. 1 Suppose t = m ， then eq. (3) can be gotten as ps follows: 1 ⎞1+ m ⎛ ps = ⎜ 1 − ⎟ ⎝ C⎠

1

en

The behavior of SS is redefined as follows. (a) All of SS should select returning number according to the global return window which BS supports； (b) Let SS sends a bandwidth request successfully after i conflict, thus SS must wait for I x m frame to participation in the new competition. (c) If the return number is greater than the number of competition window, the value of fallback state decline to the next frame after SS selects the returning number in the back window. For example, if the return number is 5 which SS selects in the current frame, and the competition windows of the current frame is 3, the competition windows of next frame is 4, thus SS sent a bandwidth request in the third transmission chance of the next frame (that is agreement with the fallback mode). The behavior of BS is the same to binary exponential return mechanism. We can predict the SS’s number nˆ which each frame participate in the competition according to the historical transmission chance utilization rate and collision rate. The return window can be set as follows:

Thus the number of bandwidth request SS which a frame sent is as follows: nr = pr n =

C≤

1 en

−1

(5)



957


be greatly reduced. In order to ensure maximum utilization rate of transmission chance with the optimal return mechanism, the initial return window and maximum return window will be set the same value, and each frame has the same number of SS in participating in the competition. The possibility of conflict is greatly improved, therefore requests discard rate also rise considerably. Then we analyze the request waiting mechanism, its initial return window and maximum return window is equal, but the request wait time is joined. That cause the number of SS is decreased, the possibility of conflicts is reduced during SS, and the bandwidth request discard rate has been reduced.

return window is equal to the maximum value of return window, which is set to the number of SS competed in. The maximum retry number of three mechanisms is set to 8. The competition windows of three mechanisms are set to 32 and 64. Two groups of simulations are obtained as follows. On the basis of the global optimal return window mechanism, request waiting mechanism joins the request waiting time, reduce the collision rate, and assure the high transmission chance utilization rate shown as Fig. 2. The transmission chance utilization rate is better than the basic mechanism and optimal return window mechanism. It is more obvious advantages when the competition for windows is 64.

Fig. 2. The utilization rate of transmission chance of three mechanisms

Fig. 3. A bandwidth request delay contrast of three mechanisms

The improved bandwidth request delay and bandwidth request loss rate of Request wait mechanism are reflected very good shown as Fig. 3 and Fig. 4 .When the competition Windows is set to 32, the maximum delay difference of optimal return mechanism and the basic mechanism is 0.7 frames. While absolute delay value is to be 90, the delay difference reaches over 5 frames. And as the number of SS increased, the delay difference of two mechanisms will be decreased gradually. But the request waiting mechanism is just the opposite; delay difference between with the basic mechanism is widened to 3.2 frames. While absolute delay value is to be 90, the delay difference is only 2.3 frames. The delay difference will increase gradually with the number of SS increased, that has greatly improved the delay of bandwidth request. This also means that the attendant packet transfer delay is also much improved. The bandwidth request discard rate of request waiting mechanism is much higher than basic mechanism. But the biggest difference is only 0.04 shown as Fig. 4. The discard rate is widened to 0.25 between request waiting mechanism and optimal return mechanism. Each SS in each frame has different return window, and the initial window is much smaller than the largest window. The return window can be expanded exponentially in case of conflict, and can largely avoid the bandwidth request conflicts. Thus the discarding rate will

Fig. 4. A bandwidth request discards delay contrast of three mechanisms

V.

Conclusion

Base on theoretical analysis of the WiMAX competition resolution mechanism, this paper describes the back window, competition window, utilization rate of transmission opportunities, discarding rate of bandwidth



958


request and the effect of bandwidth request delay, and also proposes a bandwidth competition system based on a request to wait in WiMAX system. The proposed algorithm improves the shortage of existing methods, so it has practical engineering meaning.

Authors’ information School of Engineering, Zhejiang Business Technology Institute, Ningbo 315012, Zhejing, China.

Acknowledgements This work was supported by The National Natural Science Foundation of China. (60671037).

References [1]

[2]

[3]

[4]

[5]

[6]

IEEE 802.16d-2004, IEEE Standard for Local and Metropolitan Area Networks Part 16: Air Interface for Fixed Broadband Access Systems, IEEE 802 LAN/MAN Standards Committee, 2004. IEEE 802.16e-2005, IEEE Standard for Local and metropolitan area networks-Part 16: Air Interface for Fixed and Mobile Broadband Wireless Access Systems, IEEE 802 LAN/MAN Standards Committee, 2006. E. C. Park, Efficient Uplink Bandwidth Request with Delay Regulation for Real-Time Service in Mobile WiMAX Networks, IEEE Transactions on Mobile Computing, vol. 9, n. 8, pp. 1235-1249, 2009. Namsuk Lee, Sookjin Lee, Nam Kim, A Fast Bandwidth Request Scheme in IEEE 802.16e OFDMA/TDD Systems, Proceedings of the 3st IEEE Mobile Ubiquitous Computing on Systems, Services and Technologies(Page: 173-178 Year of Publication: 2009 ISBN: 978-1-4244-5083-1 ). E. C. Park, Hwangnam Kim, Jae-Young Kim and Han-Seok Kim, Dynamic Bandwidth Request-Allocation Algorithm for Real-Time Services in IEEE 802.16 Broadband Wireless Access Networks, Proceedings of the 27st IEEE Information Computing on Computer Communications (Page: 852-860 Year of Publication: 2008 ISBN: 978-1-4244-2025-4 ). Chuck, D., Chen, K.-Y., Chang, J. M., A Comprehensive Analysis of Bandwidth Request Mechanisms in IEEE 802.16 Networks, IEEE Transactions on Vehicular Technology, vol. 59, n. 4, pp. 2046-2056, 2010.



959


An Inter-Cluster Cooperative Nodes Selection Scheme Based on Blind Channel Estimation Xiaoqiang Zhong1, Baiqing Zhou2

Abstract – An inter-cluster cooperative node selection scheme with low energy-cost is proposed for the energy –constrained wireless sensor networks. Pre-set an energy threshold first and determine a cooperative node set. Then estimate information of channels between inner-cluster cooperative nodes and cluster head nodes with a blind channel estimation algorithm that requires small amount data. The optimum inner-cluster nodes are selected as the cooperative nodes of the cluster head nodes by considering comprehensively channel states and residual energy. The blind channel estimation is then turn into the unconstrained optimization which is then resolved by quadratic programming iterative weighted method. Theoretical analysis, mathematical models, and simulation results show that the new algorithm can obtain channels information by mean of a few data. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Wireless Sensor Networks, Cooperative Communication, Blind Channel Estimation, Quadratic Programming

I.

Introduction

Constrained by size, energy and cost, nodes in Wireless Sensor Network (WSN) are required to have the minimum energy consumption which is caused by sensing module, processor module, and mainly by wireless communication module. When compared with other wireless communication systems, the WSN has a poor link and a high rate of losing package. Methods available for guaranteeing transferring data correctly in the WSN’ link layer are re-transferring data and the forward error correction [1]-[16]. However, in a poor link, only being re-transferred for many times can the data be transferred successfully, which wastes the limited bandwidth severely. While in the latter, the channel’s quality is estimated to increase effective redundant codes before sending data. Both of them not only bring enormous energy consumption, but also reduce networks’ service life. In recent years, in view of the dense nodes in WSN and the neighbor nodes around each node, cooperative communication is introduced into WSN, in which, each node combines its neighbor nodes into cluster, thus virtual Multi-input Multi-Output (MIMO) is realized. The transmission reliability is enhanced and the energy consumption is reduced as well. Based on the cooperative communication, a virtual MIMO channel is constituted by the WSN structure and receiving terminals. Cui presented in Reference [4] a single-hop cooperative MIMO transmission scheme which is based on Alamouti strategy, and built an energy consumption model that could save energy when employing virtual MIMO in WSN.


960

Reference [4]-[5] proved the energy based on the V-BLAST virtual MIMO structure, the time-lag efficiency, and the cooperative overhead effects’ on the virtual MIMO. Thereafter, a great deal of enhanced schemes is made in order to get higher energy efficiency based on the achievement. Reference [8] reviewed the cooperative virtual MIMO and proposed that two key problems in practical use, one is how to choose dynamically proper distributed cooperative nodes to form virtual antenna arrays, the other is how to modify each node’s transmission power according to the channel’s state in order to improve performance. The WSN covered in Reference [4]-[7] chose part of the neighbor nodes as cooperative nodes. Researches available took either the residual energy or the distance from nodes to the neighbors as criterion for selecting cooperative nodes in Reference [9]-[12], which studied cooperative transmission effect on energy consumption, but ignored the channel state. Reference [13] presented two criteria for selecting three cooperative nodes based on the unconstrained energy assumption. The criteria had less transmission energy consumption than that in Reference [9]-[12], but nodes’ energy with good channel would be depleted rapidly. Reference [14] covered a cooperative nodes selection scheme and made some improvement in energy efficiency when considering both the residual energy of neighbor nodes and the local channel between the leader cluster node and the neighbor nodes, but ignored the channel between the source nodes and the distal nodes. In practical use, wireless signal transfers through a time-varying multipath fading channel. In order to restore Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Xiaoqiang Zhong, Baiqing Zhou

the data received, some compensation should be made for signals passing through channel. So, in order to obtain the faded information, channel estimation is employed in the receiver. Comparing to the channel state of Source Nodes (SN) and Destination Nodes (DN), the state of Cooperative Nodes (CN) in virtual MIMO is random. A bad CN’s channel does no good for the system. It has been a key and hard problem for the WSN to estimate channel. So a channel estimation algorithm with less computation is required when considering energy consumption. And the best way is to reduce the algorithm’s demand on data quantity. Here, we propose a CN selection scheme that is based on blind channel estimation. Simulation results show that the new method could obtain accurately the channel’s information without training data.

normal distribution that has a mean of 0 and a variance of 1. (3) Channels between BS and CHNs and ICNs is aquatic-static multipath fading. Ignore that the channel coefficients satisfy the normal distribution mentioned in Reference [15]. The BS is suffered an additive white Gaussian noise. Before the cooperative communication is set up, the communication between CHNs and ICNs is a typical MIMO system. The traditional blind channel estimation that based on SOS couldn’t meet the demand of SISO system, and the HOS method is based on large quantity of data. Therefore, the key to a blind estimation method which is suitable to WSN is to select proper cooperative nodes. II.2.

II.

System Model and Inter-Cluster Nodes’ Selection Scheme II.1.

Designing Cooperative Node Selection Scheme

Considering comprehensively the state information of channel between the ICNs, their residual energy, and the Data Fusion Center (DFC), we proposed a scheme for selecting CN. The ICN with the biggest channel gain leads to a better transmission performance and a smaller energy consumption. In order to find an optimum balance between energy consumptions of nodes and transmission, Et is set as the energy threshold first when selecting CNs. Only those nodes with big residual energy (bigger than Et ) have the chance to be selected. Based on the thought from Reference [15], we design the selection scheme according to the following steps: Step 1: To choose CHNs according to the protocol of Low-Energy Adaptive Clustering Hierarchy (LEACH). Step 2: To determine sets to be selected as D(r) ( D ( r ) : {ri ∈ Dr } ), where, ri is the ith ICN. The

System Model

Usually, a WSN system model is built on the base of the WSN mode of the virtual MIMO mentioned in Reference [4], in which, channels between Cluster Head Nodes (CHNs) and Inter-Cluster Nodes (ICNs) are Single-Input Single-Output (SISO), channels between each cluster are MIMO, and channels between each cluster to Base Station (BS) are Multi-Input Single-Output (MISO), just as shown in Fig. 1.

○ Inter-Cluster Nodes (ICNs) ● Cooperative Nodes (CN)

implementation is as follows: To pre-code and pre-modulate the residual energy El ,i and ID into a short

■ Cluster Head Nodes (CHNs) MIMO Wireless Channel SISO Wireless Channel

transmission sequence which is sent to CHN through a RF front-end. The channel state information is obtained by being estimated blindly. The residual energy and the ID’s information are obtained by making convolution operation on the estimated channel and the CHNs’ sequence that is polluted by noise. When El ,i ≥ Et , the node is fell under

MISO Wireless Channel

the set to be selected (namely, D ( r ) ). If the channels of

Fig. 1. Cooperative transmission WSN model in clustering network

Let’s suppose that the BS is configured with an antenna, and could be supplied with energy, only one antenna is configured respectively for CHNs and ICNs which have constrained energy. The channel model is assumed to be: (1) Distance between each ICN is much less than that of between each CHN or between CHNs and BS. (2) Practically, CHNs’ channel is multi-path fading, so the channel coefficients satisfy the standard complex


CHNs and Node i don’t change, CHNs will feed back to the original ICNs through the original channel. Nodes that received their own ID are activated; conversely, nodes are dormant and energy-saving. Step 3: To choose the optimum node from C ( r ) , ( rk = arg max hi ), and then the Best Cooperative ri ∈C ( r )

Nodes (BCN). [1][5] Step 4: To transmit cooperatively based on the channel state information of CHNs and BCNs.


961


When

III. Blind Channel Estimation Based on Small Data Quantity

(

=

(

∑ hn si −n + ei , yi = w T xi

∑ (

i =1

(

i =1

(

)

)

)

(4)

( (

)

)

Then Equation (4) is turned into a Wolf dual one, shown in Equation (5):

)

W (α ,α ) = −ε 2

(1)

1 − 2

ε

where, the regularization parameter λ is smaller than 0. Introduce the Vapnik - ε insensitive loss function as shown in Equation (2) [16]:

ε

N

i =1

N 1 || w ||22 +λ ⋅ 1 − w T xi 2 i =1

)

)

N

minimized as shown in Equation (1):

(

(

wQP = ∑ ( ai − ai ) yi xi , γ i = λ − ai , and γi = λ − ai

= yi 2 = 1, i = 1," ,N . The cost equation is

2

(

N

+ λ ⋅ ∑ ξi + ξi − ∑ γ iξi + γiξi +

where, w , ξ , ξ are the original variables, α , α , γ ,γ are the dual variables. Work out the partial derivatives for each original variable and let them be 0, we will get the following equations:

T

1 − w T xi

N

i =1

" x N ] , xi = [ xi xi −1 " xi − M +1 ]

J (w) =

2

−∑ αi yi w T xi − 1 + ε + ξi

The equalizer’s output is a constant modulus, namely,

)

1 w 2

i =1 N

n

2

)

−∑ α i 1 − yi w T xi + ε + ξi +

The output equations of the signal received and the equalizer respectively are:

w T xi

multipliers:

L w , ξ , ξ , α , α , γ , γ =

signal received is made up of a data block with the length of N:

xi =

Lagrange

into solving saddle point of the Lagrange Function with a given λ and ε:

From the steps mentioned above, we know that on one hand, the direct way is to reduce data quantity required in the algorithm for the purpose of low energy consumption, on the other hand, a blind channel estimation with small data quantity are needed to confront the channels’ time variation. It has been a key problem to design a blind channel estimation algorithm based on small data quantity when choosing CNs in the new scheme. For convenience of study, let’s modulate CHN, CN, BCN and CFC with BPSK. The ith character of the sender sequence is Si, and si ∈ {±1} . Noise is ignored here. The

[ x1 x2

introducing

α i , α i , γ i , γi ≥ 0 , the optimization problem is then turned

(

⎧ = max ⎨0 , 1 − w T xi ⎩

)

2

⎫ −ε ⎬ ⎭

N

N

N

i =1

i =1

∑ (αi + αi ) + ∑ (αi − αi ) + (5)

N

∑∑ (αi − αi ) (α j − α j )( yi y j ) i =1 j =1

xi ,x j

In order to make the equation more compact, let: ⎡− K T Α = [α1 " α N α1 " α N ] , U = ⎢ ⎣K

(2)

K ⎤ − K ⎥⎦

where, K is a matrix constituted by:

where, ε is a non-negative number. And then introduce the slack variable factors: ξi and ξi . Rewrite the cost equation shown in Equation (1). The minimum could be obtained by solving the following optimization problems [17]:

Ki, j = yi y j xi , x j

( i, j = 1," ,N )

⋅,⋅ is an inner-product operator.

Let:

(

)

1 J w , ξ , ξ = w 2 s.t.

2

N

(

+ λ ⋅ ∑ ξi + ξi i =1

)

( w x ) −1 ≤ ε + ξ 1 − ( w x ) ≤ ε + ξ

E = - [ε − 1 ε − 1 " ε + 1 ε + 1] , T

Y = [ y1 " y N -y1 " -y N ]

T

2

T

i

i

(3)

2

T

i

The destination function is changed into a matrix function, shown in Equation (6):

i

ξi ,ξi ≥ 0 where, i = 1," , N . Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


962


1 T A UA + AT E 2 ATY = 0

min s.t.

First, pre-code and pre-modulate El ,i (residual energy of the ith ICN) and the ID into a transmission sequence with the length of 100, which is sent from ICNs’ RF front-end to CHN. The channel state is then obtained through blind estimation, shown in Fig. 2.

(6)

0 ≤ Ai ≤ λ

Up till now, the optimization problem is turned into a strict quadratic programming problem. Considering the Karush-Kuhn-Tucker (KKT) condition, the equalizer’s output is then computed by Equation (7):

Estimation Channel

N

∑ βi i =1

Channel amplitude

yk =

Practical Channel

(7)

xi , xk

where, βi are the weighted Lagrange multipliers and

satisfy βi = (αi − α i ) yi

If sequence in the sender is sent by MPSK, the ith signal is si that satisfies:

{

}

si ∈ e j 2π ( m −1 ) / M , m = 1, 2," ,P , P = 4,8," Slot time

Reconstruct the complex equalizer with:

( ) ( )

Fig. 2. Estimation of channel between the i-th ICN and CHN when modulated with OQPSK (SNR=10dB)

T

w = ⎡ Re w T ,Im w T ⎤ ⎣ ⎦ 2 M ×1 Practical Channel

and we obtain Equation (8):

Estimation Channel

⎡ Im ( xi ) ⎤ + Im ( yi ) ⎢ = ⎥ ⎣ Re ( xi ) ⎦ 2 M ×1

( (

Channel amplitude

⎡ Re ( xi ) ⎤ + g i = Re ( yi ) ⎢ ⎥ ⎣ − Im ( xi ) ⎦ 2 M ×1

(8)

) )

⎡ Re yi x*i ⎤ ⎥ =⎢ ⎢ Im y x* ⎥ i i ⎥ ⎢⎣ ⎦ 2 M ×1

where, ( ⋅) is a complex number conjugation. Then the *

(

T

constraint (namely, w xi

)

2

Slot time

) in Equation (1) can be

Fig. 3. Estimation of channel between the i-th ICN and CHN when modulated with BPSK (SNR=10dB)

. The algorithm can be reconstructed substituted by wg with the method used in BPSK signal.

IV.

Estimated signal is obtained by making convolution operations on the estimation channel and the multi-path faded and noised sequence that is received by CHN, shown in Fig. 3. A sending sequence is finally obtained, and the information about the residual energy as well. Code and modulate the node information that is qualified to join cooperative communication in the CHN, and then feed it back to the ICN through the original channel. Notice that the sequence length is shortened. For the time being, let’s assume that there is a small time delay and an unchanged channel state, when the signal passes through the original channel, the ICNs have already

Performance Simulations

(1) Usually, let the channel between one CHN and ICN satisfy the equation below: h1 = −0.0946 + 0.1715 z −1 + 0.5082 z −2 − 0.3474 z −3 +

+0.1359 z −4 − 0.0815 z −5 + 0.0130 z −6

and each node is assumed to be modulated by OQPSK. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


963


[4]

obtained the channel state information. A real sequence is obtained by making convolution operations only on the short sequence. Whether the dormancy is triggered or not depends on the sequences’ difference. The optimum cooperative nodes are then obtained by repeating the above process. (2) The original information sent is restored according to the channel estimation, shown in Fig. 4. Simulation results show that even in a low SNR (10dB), the equalizer outputs signal that matches the original sending signal with the new method that requires a few data.

[5]

[6]

[7]

[8]

[9]

[10] Channel amplitude

[11] Signal sent practically Signal estimated

[12]

[13]

[14] Signal

[15]

Fig. 4. Signal received in the CHN and signal estimated [16]

V.

Conclusion

Considering the nodes’ energy consumption and the virtual MIMO technology, we present a new blind channel estimation to obtain channel information and cooperative nodes on the premise of not increasing complicated computation and reducing the algorithm’s demand on data amount. Simulation experiments show that the new method could obtain channel information without training any data. Unfortunately, we fail to consider distance between ICNs and CHNs and to make a concrete analysis of energy consumption when setting the residual energy threshold. The blind channel estimation method covered here can be employed in other networks with constrained energy and time variant channel.


Zhejiang Business Technology Institute, Ningbo 315000, Zhejiang, China. 2

Huzhou Vocational and Technical College, Huzhou 313000, Zhejiang, China. Xiaoqiang Zhong is a teacher at Zhejiang Business Technology Institute. He thinks seriously teaching methods to improve his teaching level. His major research interest is electric automation and he has published several papers in related journals.

References [1]

[2]

[3]

Cui S, Goldsmith A J, and Bahai A, Energy-efficiency of MIMO and cooperative MIMO techniques in sensor networks, IEEE Journal Selected Areas Communication, vol. 22, n. 6, pp. 1089-1098, 2004. JayaweeraS K, Virtual MIMO-based cooperative communication for energy-constrained wireless sensor networks, IEEE Transactions on Wireless Communication, vol. 5, n. 5, pp. 984-989, 2006. Jayaweera S K, V-BLAST-based virtual MIMO for distributed wireless sensor networks, IEEE Transactions on Communications, vol. 55, n. 10, pp.1867-1872, 2007. Simi L, Berber S M, and Sowerby K W, Energy-efficiency of cooperative diversity techniques in wireless sensor networks, PIMRC’07, Athens, 2007, 9: 1-5. Nosratinia A, Hunter T E, Hedayat A, Cooperative Communication in Wireless Networks, IEEE Communications Magazine, vol. 42, n. 10, pp. 74 – 80, 2004. Bravos G N, Efthymoglou G, and Kanatas A G, MIMO-based and SISO Multihop Sensor Networks: Energy Efficiency Evaluation, WiMOB 2007, White Plains, NY, 2007, 10: 13-20. Nguyen T D, Berder O, and Sentieys O, Cooperative MIMO schemes optimal selection for wireless sensor networks, VTC2007-Spring, Dublin, 2007, 4: 85-89. Lin Z, Erkip E, and Stefanov A, Cooperative regions and partner choice in coded cooperative systems, IEEE Transactions on Communications, vol. 54, n. 7, pp. 1323-1334, 2006. Zhao B and Valenti M C, Practical relay networks: A generalization of hybrid-ARQ, IEEE Journal on Selected Areas in Communications, vol. 23, n. 1, pp. 7-18, 2005. Bletsas A, Khisti A, and Reed D P, et al., A simple cooperative diversity method based on network path selection, IEEE Journal on Selected Areas in Communications, vol. 24, n. 3, pp. 659-672, 2006. Bravos G N and Kanatas A G, Energy efficiency of MIMO-based sensor networks with a cooperative node selection algorithm, ICC2007, Glasgow, 2007, 6: 3218-3223. Zhang Yu, Cai Yueming, An Energy-efficient Adaptive Cooperative Node Selection Scheme in WSN, Journal of Electronics & Information Technology, vol. 31, n. 9, pp. 2193-2198, 2009. Zhu J, Steven C, Lyu M, Robust Regularized Kernel Regression, Systems, IEEE Trans. on Man and Cybernetics- Part B, vol. 38, n. 6, pp. 1639-1644, 2008.

Akyildiz I F, Su W, and Sankarasubramaniam Y, et al., A survey on sensor networks, IEEE Communications Magazine, vol. 40, n. 8, pp. 102-114, 2002. ZHAO J, GOVINDAN R, Understanding packet delivery performance in dense wireless sensor networks, Proceedings of International Conference on Embedded Networked Sensor Systems. New York, USA: ACM, 2003: 1-13. CAO Qing, ABDEL ZA HER T, HE Tina, et al., Cluster-based forwarding for reliable end-to-end delivery in wireless sensor networks, Proceedings of International Conference on Computer Communications. Piscataway, NJ, USA: IEEE, 2007: 1928-1936.


Baiqing Zhou received the Master degree in mechatronics engineering from Zhejiang Industry University. She is a teacher at Huzhou Vocational and Technical College. Her major research interests are computer application and image processing. She has published papers in related journals and teaching materials.


964


A Suitable Architecture and Deployment Considerations for Shallow Water Acoustic Sensor Networks V. Hadidi1, R. Javidan2, M. Keshtgary3, A. Hadidi4 Abstract – Nodes of an underwater Acoustic Sensor Network (UASN) are connected together through acoustic sound that propagates in the water better than electromagnetic waves. They are envisioned to enable applications for real-time underwater monitoring and data collection. Moreover, unmanned or autonomous underwater vehicles (UUVs, AUVs), equipped with sensors, enable the exploration of natural undersea resources and gathering of scientific data in collaborative monitoring missions. In this paper, several fundamental key aspects of underwater acoustic communication are investigated, Different architectures for two-dimensional and three-dimensional underwater sensor networks are discussed, and the main challenges for the development of efficient networking solutions are presented. Next, we propose a new multipath scheme for shallow water, which can guarantee certain end-to-end packet error rate while achieving a good balance between the overall energy efficiency and the end-to-end packet delay. Finally an appropriate localization algorithm for shallow waters is proposed. The simulations results on prototype data show the effectiveness of the proposed architecture and localization algorithm. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Acoustic Communications, Routing Protocols, Architecture of Underwater Acoustic Sensor Networks, Localization Algorithm

I.

Optical waves do not suffer from such high attenuation but are affected by scattering. Thus, links in the underwater networks are usually based on acoustic wireless communications[1],[4].Even there exist many recently developed network protocols for wireless sensor networks, the unique characteristics of the underwater acoustic communication channel require very efficient and reliable new data communication protocols [5]. Major challenges in the underwater acoustic networks are: 1. Propagation delay is five orders of magnitude higher than in radio frequency terrestrial channels and is also variable; 2. The underwater channel is severely impaired, especially due to multipath and fading problems; 3. The available bandwidth is severely limited; 4. High bit error rates and temporary losses of connectivity (shadow zones) can be experienced; 5. Sensors may fail because of fouling and corrosion; 6. Battery power is limited and usually batteries cannot be easily recharged, also because solar energy cannot be exploited. 7. In the shallow waters, there are many problems regarding multipath and seabed and surface scattering. Recent advances in the research and development of underwater wireless sensor networks have produced much novel architecture and protocols, most of these methods are suitable for deep water. But they have poor

Introduction

Sensor networks are a very challenging problem. One of the earliest sensor networks prototypes was Sound Surveillance System (SSS), which was a system of acoustic sensors (hydrophones) on the ocean bottom and deployed at strategic locations to detect and track quiet submarines [1]. Such sensor networks were expensive, fixed, and cable-connected. Although they were too expensive to be used in many applications, they gave original ideas about underwater sensor networks. With the rapid progress in technology, the situation is changed [2]. The new network technologies focus on distributed information processing, efficient power consumption, and etc. Most of these researches are about ground sensor networks, such as Unattended Ground Sensors [3] and only a few researches were related to underwater sensor networks. Underwater environment is very different from the open ground environment. The pressure, humidity, and changing environment are the main restrictions against the design of the underwater sensor networks. Therefore, integrating the recent advanced technology with the underwater sensor networks should be considered carefully. Acoustic communications are the typical physical layer technology in the underwater networks. In fact, radio waves propagate through conductive sea water only at extra low frequencies (30 - 300 Hz), which require large antennae and high transmission power. Manuscript received and revised October 2011, accepted November 2011

965


V. Hadidi, R. Javidan, M. Keshtgary, A. Hadidi

meters may have more than a hundred kHz bandwidth. In both cases these factors lead to low bit rates [6]. Moreover, the communication range is dramatically reduced as compared to the terrestrial radio channel. Underwater acoustic communication links can be classified according to their range as very long, long, medium, short, and very short links [7]. Acoustic links are also roughly classified as vertical and horizontal, according to the direction of the sound ray. Acoustic communications in shallow water encountered a problem due to the channel characteristics of the underwater acoustic channel. For long range underwater acoustic communications, the main problem is the presence of multipath propagation caused by reflection and scattering of the transmitted signals at the bottom and the surface. Reflections from channel boundaries and diverse objects dominate the multipath structure. The transmitted signal can go through multiple paths in order to reach the receiver. These multiple paths can cause significant time spread in received signal. Each path has can possibly have multiple surface interactions causing additional frequency spreading due to motion of the water. Shallow water propagation is very sensitive to changes in the geometrical parameters like water depth, source-receiver range or bottom slope leading to variations in the impulse response of the underwater acoustic sound channel. Normal mode approaches have been widely used in underwater acoustics and are derived from an integral representation of the wave equation. When propagation is described in terms of normal modes, changes in the environment translate into energy transfer between modes. In the following, we analyze the factors that influence acoustic communications in order to state the challenges posed by the shallow water channels for underwater sensor networking. These include: Path loss: Attenuation is mainly provoked by absorption due to conversion of acoustic energy into heat, which increases with distance and frequency. It is also caused by scattering and reverberation (on rough ocean surface and bottom), refraction, and dispersion (due to the displacement of the reflection point caused by wind on the surface). Water depth plays a key role in determining the attenuation. Geometric Spreading refers to the spreading of sound energy as a result of the expansion of the wave fronts. It increases with the propagation distance and is independent of frequency. There are two common kinds of geometric spreading: spherical (Omni-directional point source), and cylindrical (horizontal radiation only). Noise: Man made noise. This is mainly caused by machinery noise (pumps, reduction gears, power plants, etc.), and shipping activity (hull fouling, animal life on hull, cavitations).

performance in the shallow water where, severe signal degradation can occur in such a channel due to multipath effects and the refractive properties of the channel, which may include multiple interactions with the sea bottom and sea surface. Time-varying multipath propagation is one of the major factors that limit acoustic communication performance in the shallow water. The QoS of such networks is limited by the low bandwidth of acoustic transmission channels, high latency resulting from the slow propagation of sound, and elevated noise levels in some environments. The long-term goal in the design of underwater acoustic networks is to provide a self-configuring network of distributed nodes with network links that automatically adapt to the environment through selection of the optimum system parameters. Multipath schemes are commonly believed to be beneficial to load balance and network robustness, but they are usually not considered energy-efficient since more nodes will be involved in a multipath scheme than in a one-path scheme. In this paper, contrary to the common intuition, we show that in underwater fading environments, for timecritical applications, if multipath schemes are properly combined with power control at the physical layer and packet combining at the destination, significant energy savings can be achieved with low end-to-end delays. Furthermore we propose localization algorithm for shallow water environment, which is applicable to a number of applications, including underwater target tracking, seismic monitoring, equipment monitoring, leak detection, etc. The simulations results on prototype data show the effectiveness of the proposed architecture and localization algorithm. The rest of this paper is organized as follows: In Section II basic of acoustic communications is discussed. Section III reviews the related works on underwater sensor networks. The general architectures of underwater sensor networks and routing protocols are the subject of Section IV. In Section V, a new method for shallow water sensor networks architecture is presented and it was shown that it can be an ideal localization algorithm without errors for underwater usage. In addition, the simulation results and analyzes of the output using different parameters are demonstrated. Finally, conclusions and remarks are outlined in Section VI.

II.

Basic of Acoustic Communications

Underwater acoustic communications are mainly influenced by path loss, noise, multi-path, Doppler spread, high and variable propagation delay. All these factors determine the temporal and spatial variability of the acoustic channel, and make the available bandwidth of the Under Water Acoustic (UW-A) channel limited and dramatically dependent on both range and frequency. Long-range systems that operate over several tens of kilometers may have a bandwidth of only a few kHz, while a short-range system operating over several tens of



966


deployment grid structures may not be feasible for the underwater environment. In particular, in [8], methods for determining network connectivity and coverage given a node-reliability model are discussed, and an estimate of the minimum required node-reliability for meeting a system-reliability objective is provided. An interesting result is that connectivity does not necessarily imply coverage. As the node reliability decreases, in fact, the sufficient condition for connectivity becomes weaker than the necessary condition for coverage. Although [8] provides useful theoretical bounds and insight into the deployment of wireless terrestrial sensor networks, the analysis is limited to grid structures. In [9], two coordination sleep algorithms are compared, a random and a coordinated sleep scheme. It is shown that when the density of the network increases, the duty cycle of the network can be decreased for a fixed coverage. Although [9] provides sound coverage algorithms for terrestrial sensor networks, its results cannot be directly applied to the underwater environment where the sensor density is much lower than in the terrestrial case, and spatiotemporal correlation cannot often be assumed [10]. In [11], sensor coverage is achieved by moving sensor nodes after an initial random deployment. However, [11] requires either mobile sensor nodes or redeployment of nodes, which may not be feasible for UWASNs. In [12], sensing and communication coverage in a three dimensional environment are rigorously investigated. The diameter, minimum and maximum degree of the reach ability graph that describes the network are derived as a function of the communication range, while different degrees of coverage (1-coverage and, more in general, k-coverage) for the 3D environment are characterized as a function of the sensing range. Interestingly, it is shown that the sensing range required for 1-coverage is greater than the transmission range that guarantees network connectivity. Since in typical applications t ≥ r, the network is guaranteed to be connected when 1-coverage is achieved. Although these results were derived for terrestrial networks, they can also be applied in the underwater environment. Thus, in this paper, we will focus on the sensing coverage when discussing deployment issues in 3D UWASNS, as in three-dimensional networks it implicitly implies the communication coverage. A recent work combines short base line (SBL) long base line (LBL) by using short-range Global Positioning System (GPS)enabled stationary buoys for autonomous underwater vehicle (AUV) tracking applications [13]. Although SBL and LBL can be utilized for localization of disconnected individual underwater equipment, they are not convenient for UASNs. SBL requires the operation of a ship which is costly and unscalable for UASNs, whereas the long-range signals of LBL have the possibility of interfering with the communication among UASN nodes. Among the proposed solutions, GPS-based localization

Ambient Noise is related to hydrodynamics (movement of water including tides, currents, storms, wind, rain, etc.), seismic and biological phenomena. Multi-path: Multi-path propagation may be responsible for severe degradation of the acoustic communication signal, since it generates Inter-Symbol Interference (ISI). The multi-path geometry depends on the link configuration. Vertical channels are characterized by little time dispersion, whereas horizontal channels may have extremely long multi-path spreads, whose value depend on the water depth. High delay and delay variance: The propagation speed in the underwater acoustic (UWA) channel is five orders of magnitude lower than in the radio channel. This large propagation delay (0.67 s=k m) can reduce the throughput of the system considerably. The very high delay variance is even more harmful for efficient protocol design, as it prevents from accurately estimating the round trip time (RTT), key measure for many common communication protocols. Doppler spread: The Doppler frequency spread can be significant in UWA channels [7], causing a degradation in the performance of digital communications. Transmissions at a high data rate can cause many adjacent symbols to interfere at the receiver, requiring sophisticated signal processing to deal with the generated ISI. Most of the described factors are caused by the chemical physical properties of the water medium such as temperature, salinity and density, and by their spatiotemporal variations. These variations, together with the wave guide nature of the channel, could cause the acoustic channel to be temporally and spatially variable. In particular, the horizontal channel is by far more rapidly varying than the vertical channel, in both deep and shallow water.

III. Background and Related Work Past several years have witnessed a rapidly growing interest in UWSNs from both academia and industry. Many applications, networking protocols and devices have been introduced. However, most of them are application specific, and usually lack compatibility with each other. Moreover, due to limited resources, majority of work on UWSNs remains in the stage of computer simulation. Further, with different assumptions and platforms, it is very difficult to compare solutions for similar problems. Therefore, it is imperative to have a generic architecture to facilitate UWSN research. The problem of sensing and communication coverage for terrestrial sensor networks has been addresses in several papers. Many previous deployment solutions and theoretical bounds assuming spatiotemporal correlation, mobile sensors, redeployment of nodes, and particular



967


schemes are not suitable for UASNs since the highfrequency GPS signals do not propagate well in water, whereas GPS-less schemes are generally not convenient since they require large amounts of messaging between sensor nodes. In [14] the authors propose a localization scheme that uses an AUV. In the AUV-Aided Localization (AAL) scheme, underwater sensor nodes are stationary, and an AUV travels underwater to localize the sensor nodes. The AUV periodically surfaces to receive GPS coordinates, and does dead-reckoning for tracking selflocation while submerged. The AUV broadcasts wake-up messages from different places on its route, and the underwater sensor nodes start the localization process upon hearing these messages. In [15], the authors propose a distributed hierarchical localization scheme for stationary UASNs. The hierarchical architecture of Large Scale Localization (LSL) employs three types of nodes: surface buoys, anchor nodes, and ordinary sensor nodes. Surface buoys float on the surface and learn their coordinates through GPS. Anchor nodes and ordinary sensor nodes float underwater. Anchor nodes are assumed to be localized by the surface buoys at an earlier deployment stage, and LSL considers only the localization of ordinary sensor nodes. In the ordinary sensor localization process, anchor nodes periodically broadcast their coordinates, while ordinary nodes send short messages periodically to measure distances to their neighbors via Time-of-Arrival (ToA). LSL has a hierarchical structure, which means this scheme can be used in large-scale UASNs. Its main drawback is having high energy consumption and overhead due to beacon exchanges, localization messages, and the messages forwarded by unlocalized nodes. In [16], the authors utilize the same hierarchical architecture of [15] and propose Scalable Localization with Mobility Prediction (SLMP) for mobile UASNs. Anchor nodes and ordinary nodes estimate their locations by using their previous coordinates and their mobility patterns. In a mobile UASN, mobility patterns may become obsolete in time; therefore, anchor nodes periodically check the validity of their mobility pattern and trigger an update when necessary. An anchor node, after predicting its location, uses surface buoy coordinates and distance measurements to buoys to estimate its location. Underwater objects are moving continuously with water currents and dispersion. Researches in hydrodynamics show that the movement of underwater objects is closely related to many environment factors such as the water current, water temperature [17], [18]. In different environments, the mobility characteristics of underwater objects are different. For example, the mobility patterns of objects near the seashore demonstrate a certain semi periodic property because of tides; but for objects in rivers, their moving behaviors have no such property.

IV.

Architecture of Underwater Acoustic Sensor Networks [1] IV.1. Static Architecture

For static architecture, the network topology would be in relative static state after sensors were deployed, the network could be anchored into two-dimensional (2D) or three-dimensional (3D) either on the sea floor or surface. The main character of this architecture is that the topology doesn’t change or move after deployment. In 2D case, the topology could be grid, cluster, tree, or linerelay deployment same as terrestrial wireless sensor networks (WSNs). In 3D case, sensors could be moored to anchors on ocean floor or to surface floats with fix depth. Fig. 1 illustrates the clustered deployment in 2D space, either on the surface or seabed. And Fig. 2 illustrates the 3D case, especially in shallow water. 1) Two-dimensional case A reference architecture for two-dimensional underwater networks is shown in Fig. 1. A group of sensor nodes are anchored to the bottom of the ocean with deep ocean anchors [19]. Underwater sensor nodes are interconnected to one or more underwater sinks (uwsinks) by means of wireless acoustic links. Uw-sinks, as shown in Fig. 1, are network devices in charge of relaying data from the ocean bottom network to a surface station. To achieve this objective, uw-sinks are equipped with two acoustic transceivers, namely a vertical and a horizontal transceiver. The horizontal transceiver is used by the uw-sink to communicate with the sensor node s in order to: (i) send commands and configuration data to the sensors (uw-sink to sensors); (ii) collect monitored data (sensors to uw-sink). The vertical link is used by the uwsinks to relay data to a surface station. In deep water applications, vertical transceivers must be long range transceivers as the ocean can be as deep as 10 km. The surface station is equipped with an acoustic transceiver that is able to handle multiple parallel communications with the deployed uw-sinks [20],[21]. It is also endowed with a long range RF and/ or satellite transmitter to communicate with the onshore sink (os-sink) and /or to a surface sink (s-sink).

Fig. 1. Architecture for 2D underwater sensor networks [19]



968


network reconfiguration, or even a normal sensor for data sensing. Fig. 3 illustrates a general architecture used in ocean surveillance system. The mobile nodes can be ROVs or AUVs, which equipped with two transceivers, in horizontal and vertical direction. With the help of mobile nodes, the network can realize 3D surveillance, and adjust the topology to meet the application demand. Mobile nodes can also optimize the deployment to maximize network capacity, and ensure the reliability when some sensors lose their function.

2) Three-dimensional case Three dimensional underwater networks are used to detect and observe phenomena that cannot be adequately observed by means of ocean bottom sensor nodes, i.e., to perform cooperative sampling of the 3D ocean environment [22]. In three-dimensional underwater networks, sensor nodes are floated at different depths in order to observe a given phenomenon [23]. One possible solution would be to attach each uw-sensor node to a surface buoy, by means of wires whose length can be regulated so as to adjust the depth of each sensor node [24]. However, although this solution allows easy and quick deployment of the sensor network, multiple floating buoys may obstruct ships navigating on the surface, or they can be easily detected and deactivated by enemies in military settings [25]. Furthermore, floating buoys are vulnerable to weather and tampering or pilfering. For these reasons, a different approach can be to anchor sensor devices to the bottom of the ocean. In this architecture, depicted in Fig. 2, each sensor is anchored to the ocean bottom and equipped with a floating buoy that can be inflated by a pump. The buoy pushes the sensor towards the ocean surface. The depth of the sensor can then be regulated by adjusting the length of the wire that connects the sensor to the anchor, by means of an electronically controlled engine that resides on the sensor [19]. A challenge to be addressed in such architecture is the effect of ocean currents on the described mechanism to regulate the depth of the sensors.

Fig. 3. Hybrid architecture: static sensor with mobile nodes [22]

Inhybrid network architecture was discussed in [26], there are four kinds of nodes to obtain a tiered deployment. At the lowest layer, large numbers of sensors are deployed on the seabed for data collection. One or more control nodes with connections to Internet are deployed on ocean surface or off-shore platform. Another two types are super nodes which can interconnect with high speed network and relay data efficiently, and submersible robots. A network topology using Delay-tolerant Data Dolphins (DDD) in stationary sensor grids was described in [27], DDDs with high energy can help to maximize the network lifetime of UASNs. Each underwater sensor is only required to transmit its data to the nearest dolphin within one-hop distance as dolphins’ approaching. An architecture for short-term time-critical aquatic exploration applications was described in [28], using UASN to control ROV remotely for emergency underwater investigation .A three layers architecture was considered in [29], different physical environments present different requirements for sensors. The fixed seafloor sensors require high robustness for long term data collection, the surface nodes acting as sinks are fixed in special regions and with GPS for localization. The mobile nodes between upper two layers are AUVs or ROVs, which can move horizontally and vertically. We designed an application model for underwater monitoring applications shown in Fig. 4 [30]. It works in two ways: local base station collects sensors’ data with regular time resolution, and mobile actors collect data from virtual clusters with high temporal resolution. Mobile actors dive into local monitoring area from surface ships or submarines and then scatter to different

Fig. 2. Architecture for 3D underwater sensor networks [19]

IV.2. Hybrid Architecture We call this as Hybrid Architecture where an UASN consists of lots of static sensors together with some mobile sensors, no matter AUVs, ROVs or any other sea gliders. In hybrid architecture, mobile nodes play a key role for additional support in accomplishing task, perhaps for data harvesting or enhancing the network capacity. Mobile nodes could be considered as super nodes which has more energy and can move independently, and it could be a router between fixed sensors, or a manager for



969


[28]. Low-cost sensors are deployed to cover a special continuous monitoring region, data are collected by local sensors, be relayed by intermediate nodes, and finally reach the gateway on the surface level which equipped with both acoustic and RF transceiver. Based on this architecture, a surface-level gateway deployment method was given in [31]. All the surface gateways form virtual sinks want to find the tradeoff between the number of surface gateways and the expected delay and energy consumption. A mobility model for coastal underwater environments was presented in [32], free-floating sensors are initially deployed in a small sub area and would be shifted by the effect of meandering sub-surface currents and vortexes in a large coastal environment. Another class of ocean monitoring networks was proposed in [33], free-floating underwater devices operate autonomously and collaborate through an acoustic communication among them, the drogue devices drift freely with the ocean currents and equip with a buoyancy control piston. The important application is for short term pollutant tracking. A majority of this kind of architecture is passive , the topology would be changed according to the integrated effect of currents, winds, tides, and waves. This brings the biggest disadvantage that the coverage and communication link could not be guaranteed, and it is difficult to achieve effective topology control. The interesting characteristic of this architecture is that it can track objects moving with water currents without any manual interference. For the scenarios that mobile UASN works together with free floating sensor network, communication connection is needed between under water layer and surface layer, this calls for some free floating sensors on the surface should be equipped with acoustic modem, also under water autonomous mobile nodes is needed for keeping active data link.

regions. As mobile actors approach, sensors would selforganize to form temporary clusters according to the number of actors. Each actor serves as cluster head and collects data from all nodes within that cluster by multihops. After data was collected, it would switch back to local networks. Because the clusters are alive just for a short period, they are called virtual clusters. We proposed a new strategy which is composed of three algorithms: Area partition and scattering, sub-region optimizing and virtual cluster formation algorithm. Experiment results indicate that our strategy can reduce sensors’ power consumption significantly and achieve lower end-to-end delay.

Fig. 4. UASNs’ application model with multiple mobile actors [22]

We can consider the hybrid architecture as a special mobile ad-hoc network, in which the mobile nodes are not only task operators, but also network managers. Data relayed by mobile nodes can shorten the end-to-end delay. On the other hand, it can prolong the lifetime of static network. The mobile nodes build a link between surface level and seabed level, and the ability of the mobile nodes decides the efficiency of the network. Its shortcoming is that it would be expensive and difficult for practical applications. IV.3. Mobile UASNs and Free Flouting Networks In this architecture, all the nodes are not restricted geographically, nodes can move freely, and the network topology would be variable. Fig. 5 shows a normal architecture that could be divided into two layers, surface layer sensors and underwater layer. Surface layer sensors often equipped with a wireless transceiver for data communication and are usually used for pollution detecting, water quality surveillance, coastal circulation monitoring and pollutant tracking. Underwater layer consists of many mobile nodes which can stay at any depth with the help of float equipment; it can be used for ocean biogeocenose investigation, fish migration and biological monitoring. Unlike the active behavior of mobile nodes, the sensors movement is passive; the dynamic ocean characters (such as waves, tides, currents) play a key role on the nodes movement. A mobile UASN architecture for long-term non timecritical aquatic monitoring applications was proposed in

Fig. 5. Mobile UASNs with free-floating networks [22]

IV.4. Sensor Networks with Autonomous Underwater Vehicles AUVs can function without tethers, cables, or remote control, and therefore they have a multitude of applications in oceanography, environmental monitoring, and underwater resource studies. Previous experimental



970


routes for the first time and each time the network topology is modified because of mobility, node failures, or channel state changes because updated topology information must be propagated to all network devices. In this protocol, each device can establish a path to any other node in the network, which may not be required in underwater networks. Also, scalability is an important issue for this family of routing schemes. For these reasons, proactive protocols may not be suitable for underwater networks. Reactive protocols (e.g., ad hoc on-demand distance vector [AODV], dynamic source routing [DSR]) are more appropriate for dynamic environments but incur a higher latency and still require source-initiated flooding of control packets to establish paths. Reactive protocols may be unsuitable for underwater networks because they also cause a high latency in the establishment of paths, which is amplified underwater by the slow propagation of acoustic signals. Geographical routing protocols (e.g., greedy facegreedy [GFG], partial-topology knowledge forwarding [PTKF]) are very promising for their scalability feature and limited signaling requirements. However, global positioning system (GPS) radio receivers do not work properly in the underwater environment. Still, underwater sensing devices must estimate their current position, irrespective of the chosen routing approach, to associate the sampled data with their 3D position. In [35], the authors propose a localization scheme for underwater sensor networks that reduces the 3D localization problem to a 2D problem by a nondegenerative projection technique preserving network localizability. The scheme is based on the fact that there should be at least d + 1 anchor nodes to uniquely localize a network in d dimensions. Projections of these anchors are taken in a 2D plane containing the node to be localized. Based on these projections and its own depth information, the node can localize itself successfully if the x and y co-ordinates of the anchor are distinct. Some recent work proposed network-layer protocols specifically tailored for underwater wireless networks. In [36], a long-term monitoring platform for underwater sensor networks consisting of static and mobile nodes, also called mules, is proposed, and hardware and software architectures are described. The nodes communicate point-to-point, using a high-speed shortrange optical communication system, and broadcast using an acoustic protocol. The mobile nodes can locate and hover above the static nodes for data mulling and can perform useful network maintenance functions such as deployment, relocation, and recovery. However, due to the limitations of optical transmissions, communication is enabled only when the sensors and the mules are in close proximity. In [37], the authors propose a geographical routing protocol that favors paths with minimal amount of “zigzagging” and that can find all possible paths to reach the destination. Initially, data packets are routed with minimum energy in a cone-shaped region whose axis

work has shown the feasibility of relatively inexpensive AUV submarines equipped with multiple underwater sensors that can reach any depth in the ocean. The integration of UW-ASNs with AUVs requires new network coordination algorithms such as: Adaptive sampling: This includes control strategies to command the mobile vehicles to places where their data will be most useful. For example, the density of sensor nodes can be adaptively increased in a given area when a higher sampling rate is needed for a given monitored phenomenon. Self-configuration: This includes control procedures to automatically detect connectivity holes due to node failures or channel impairment, and request the intervention of an AUV. Furthermore, AUVs can either be used for installation and maintenance of the sensor network infrastructure or to deploy new sensors. One of the design objectives of AUVs is to make them rely on local intelligence and be less dependent on communications from online shores. In general, control strategies are needed for autonomous coordination, obstacle avoidance, and steering strategies. Solar energy systems allow increasing the lifetime of AUVs, i.e., it is not necessary to recover and recharge the vehicle on a daily basis. Hence, solar powered AUVs can acquire continuous information for periods of time of the order of months. A reference architecture for 3D UW-ASNs with AUVs is shown in Fig. 6.

Fig. 6. 3D Underwater Sensor Networks with AUVs [23]

IV.5. Routing Protocols There are several drawbacks with respect to the suitability of the existing terrestrial routing solutions for underwater wireless communications. Routing protocols can be divided into three categories, namely, proactive, reactive, and geographical. Proactive protocols (e.g., destination sequenced distance vector [DSDV], optimized link state routing [OLSR]) provoke a large signaling overhead to establish



971


acoustic and RF modems are strategically deployed at the water surface. Each underwater sensor node can monitor and detect environmental events locally as shown in Fig. 7. When an underwater sensor node has data to report, it first transfers the data toward one or multiple surface gateway nodes (each is also referred to as a sink) through acoustic links. Then, these surface gateway nodes relay the received data to the control center through radio links. Compared to the acoustic links in water, surface radio links are much more reliable, faster, and more energy-efficient. Considering that radio signal propagation is orders of magnitude faster than acoustic signal propagation, it is safe to assume that surface gateways can send packets to the control center in negligible time and with relatively small energy consumption (acoustic communications consume much more energy than radio communications. In addition, gateway nodes are usually more powerful and have more energy supplies.). In this way, all the surface gateways (or sinks) form a virtual sink. This multisink network architecture is helpful in traffic balance and multiple-path finding, as has been studied and analyzed in [41], [42], and [43]. For our scheme, this multisink architecture can effectively help to find more paths to the (virtual) sink (since any surface gateway is counted as a sink) and can greatly reduce the packet-collision probability in the MAC layer.

passes through the sender and the receiver. The transmission power is increased until an intermediate relay node is found. If there are no nodes in that region, the axis of the cone is shifted until the packet is forwarded to a relay node. In [38], two distributed routing algorithms for delayinsensitive and delay-sensitive applications are introduced, which allow each node to select the optimal next hop, transmit power, and strength of the forward error correction algorithm. Their objective is to minimize the energy consumption, while taking the condition of the underwater acoustic channel and different application requirements into account. Analysis of multihop versus single-hop routing solutions is performed in [39], [40]. In [39], the authors investigate the delay-reliability trade-off for multihop underwater acoustic networks and compare multihop versus single-hop routing strategies while considering the overall throughput. The analysis shows that increasing the number of hops improves both the achievable information rate and reliability, which captures the decay rate of the decoding error probability because the coding block length increases asymptotically. In [40], a multihop underwater acoustic network is analyzed to understand the effect of frequency and reuse factor on signal-to-noise-ratio (SNR) and interference strength. Based on numerical analysis, the article concludes that most of the interference at the sender is contributed by the two or three nearest interfering nodes. Although the analysis provides interesting insights, it relies on the limiting assumption that nodes are arranged in a line.

V.

The Proposed Methods

V.1.

Architecture Network Model

In an underwater sensor network, with high probability, multipath routing protocols can find multiple paths between any two nodes because of the relatively high node density. This assumption holds even stronger in the multiple-sink underwater network architecture. Different paths will experience independent fading if they are node-disjoint. This work utilizes this property to provide “multipath macro-diversity”. Specifically in this technique, the source node transmits the same packet along multiple paths to the same destination. The transmission power at each intermediate node along each path is controlled by the source nodes based on the path characteristics. Multiple copies of the packet (some of these copies may be corrupted during transmission) will arrive at the destination along different paths, and the destination then recovers the packet by combining the received copies. We consider the following multisink underwater sensor network model: Underwater sensor nodes with acoustic modems are densely distributed in a 3-D aqueous space, and multiple gateway nodes with both

Fig. 7. Network model

As shown in Fig. 8, first, the source node (any underwater sensor node in our network model can be a source node) initiates a multipath routing process to find paths from the source to the destination (in our network model, the control center can be the destination). Through this route-finding process, the source will get to know some network parameters such as path length and the number of available paths. Based on this knowledge, the source node selects some paths and calculates the optimal transmitting power for each node along the selected paths. Then, it sends the same packet along the selected paths. Intermediate nodes on these selected paths will relay the packet with specified transmitting power



972


retransmit the packet. We set the maximal times of retransmission nr . After retransmitting a packet for nr times, a node will simply drop this packet.

parameters (carried in the packet header). When the destination receives all copies of the packet (some copies may get corrupted), it performs packet combining to recover the original packet.

B. Results and Analysis Fig. 9(a) shows the impact of end-to-end packet error rate(PER) on various schemes. From this figure, we can observe that with the increase of end-to-end PER, the average energy consumption per packet will decrease sharply. Compared to one-path without retransmission, our method always consumes much less energy. Fig. 9(b) clearly shows that our scheme can achieve high energy efficiency with small end-to-end delay under certain end-to-end PER requirements. Fig. 9(c) also shows that the end-to-end delivery ratio is almost the same for these three schemes.

Fig. 8. Basic procedure of multipath routing

V.2.

Performance Evaluation of Architecture Model

A. Simulation Settings Following the multiple-sink underwater sensor network model, the simulated network settings are as follows: 512 underwater sensor nodes are randomly deployed in a three-dimensional space of 4000×4000×2000 m3; 36 surface gateways are deployed in a two-dimensional area of 4000×4000 m2 at the water surface. A node can use any surface gateway as long as it can find a path to the gateway, and the node is not required to send its packets to all gateways. Unless specified otherwise, the simulation parameters are as follows: The maximal transmission range of underwater sensor nodes is set to 600 m, and the data rate is set to be 10 kb/s. Each simulation lasts for 10000 s. Thus, each node generates about 1000 packets in each simulation. We run simulations for 100 times and take the average as our final results. For comparison, we implement two other schemes in the same underwater network settings. One scheme is one-path transmission with power control but without retransmission (referred to as one-path without retransmission for short), and the other scheme is onepath transmission with retransmission and power control (referred to as one-path with retransmission for short). In the one-path without retransmission scheme, through a routing process, the source node first finds the most energy-efficient path and transmits its packets with power control to guarantee the end-to-end packet error rate. No retransmission is performed upon transmission failure. For the one-path with retransmission scheme, it works as follows: First, the source node finds the most energy-efficient path by its routing process, and then transmits packets with power control along this path. Retransmission is allowed upon failure (i.e., if the sender does not receive an ACK for a packet from the receiver after time tr (in our simulations, we set tr=1 s), it will

(a) Average energy consumption per packet

(b) Average end-to-end packet delay

(c) End-to-end delivery ratio Figs. 9. Performance with varying end-to-end PER

This is because all of them are designed to adjust the



973


incur non-trivial extra cost. On the other hand, the Timeof-Arrival (ToA) algorithm can be used in underwater environments measuring arrival time by using acoustic signal only. Once the distance estimation is made, we use general position computation techniques which are called "Trilateration" and "Multilateration". Therefore general localization algorithm in UWSNs is constituted with three steps. 1st Step is distance estimation using ToA technique. 2nd Step is position computation using Trilateration and Multilateration techniques. 3rd Step is determination of sensor position[57], [58].

transmitting power of nodes to minimize the overall energy consumption with certain end-to-end packet error rate. As shown in Fig. 9(c), all these three schemes can achieve the desired reliability for data packet delivery well. Multipath schemes are commonly believed to be beneficial to load balance and network robustness [44], [45], [46], but they are usually not considered energyefficient since more nodes will be involved in a multipath scheme than in a one-path scheme. This approach, contrary to the common intuition, shows that in underwater fading environments, for time-critical applications, if multipath schemes are properly combined with power control at the physical layer and packetcombining at the destination, significant energy savings can be achieved with low end-to-end delays [47], [48], [49]. V.3.

Localization Algorithm

As depicted at Fig. 10, the localization algorithm can be divided into three distinct steps: - Distance/ Angle estimation - Position computation - Localization algorithm Several number of reference nodes (RNs) are often called "anchors" or "base stations", since their positions are known. Every possible point in the area has to be covered by at least three RNs. In this way any point can be localized. Unfortunately this method is very impractical since it is not always possible to cover all area with only three RNs. On the other hand, the use of a larger number of RNs increases overall costs. Thus placing additional RNs may be impossible [50], [51]. So what we propose is to transfer reference node function to every sensor node. Because sensors can communicate with each other, they can also estimate their distances each other. In this way, it is possible to localize one sensor using other sensors with known position. To realize underwater applications, we can borrow many design principle and tools from ongoing groundbased research. Unfortunately, we only have a few choices in UWSNs because the nature of UWSNs is fundamentally different from the terrestrial sensor networks [52], [53], [54], [55]. Because RF signals cannot travel far in the underwater due to severe absorption losses, acoustic communications is used instead of RF signal. Commonly used distance estimation techniques are four kinds of methods in terrestrial sensor networks. First, the Received-Signal-Strength (RSS) algorithm is vulnerable to acoustic interferences, such as near-shore tide noise, near-surface ship noise, multi-path, and Doppler frequency spread, etc. Second, the TimeDifference-of- Arrival (TDoA) algorithm which uses RF and acoustic signal is no longer feasible as the RF signal fails in underwater[56]. Third, the Angle-of-Arrival (AoA) algorithm requires directional transmission/reception devices, which would

Fig. 10. Localization procedures

Distance estimation: The ToA is most simple and intuitive among distance estimation techniques. This algorithm estimates the distance between nodes by measuring the propagation time of a signal. So it requires precise time synchronization between two nodes. In this case the distance between two nodes is directly proportional to the time the signal takes to propagate from one point to another. This way, if a signal is sent at time t1and reached the receiver node at time t2, the distance between two nodes can be defined as:

d = sr ( t2 − t1 )

(1)

where sr is the propagation speed of acoustic signal (1500 m/s).The time when the signal leaves the node must be in the packet. Position computation: The trilateration technique is the most basic and intuitive position computation method. This method computes the position of sensor node via the intersection of three circles, as depicted in Fig. 11. To estimate its position using the trilateration technique, a node needs to know the positions of three reference nodes and its distance from each of these nodes. Distances can be estimated using the ToA. Furthermore, when a larger number of reference points are available, we can use the multilateration



974


where,(xi,yi) is the position of the ithactual sensor node, and (x*i,y*i) is the position of the ithestimated sensor node with errors. And j indicates the number of the sensor nodes considered[60].

technique to compute the position of sensor node. In this case, an over determined system of equations must be solved: D1 =

( x1 − xm )2 + ( y1 − ym )2

D2 =

( x2 − xm )2 + ( y2 − ym )2

D3 =

( x3 − xm )2 + ( y3 − ym )2

V.4.

(2)

Performance Evaluation of Localization Algorithm

A. Simulation Settings It is assumed that nodes are randomly deployed into 1000×1000 m2 field. Every sensor node can communicate with other sensors if the distance between them is smaller than sensor communication range R. And initial RNs are placed on the edge of a field and its parameter R=1000. Fig. 12 shows the initial random sensor distribution.

where (xm,ym) is the position of unknown node. In the position computation step, there are two exceptions in order to select reference nodes. In these two cases, it is not possible to compute unknown node: - When two already known nodes have the same position - When all three nodes lie on one line.

Fig. 11. Position computation methods

Localization algorithm: In UWSNs, thousands of sensors are often deployed in hostile environments, i.e. battlefield, endangered or security areas. Sensors are deployed using ship or artillery. So we can assume that their positions are random and unknown. In the 1st step, from the random sensor distribution, we perform the localization for unknown nodes which are in overlap area of large range initial RNs. At this stage, we consider only three initial RNs and if we determine the positions of all sensor nodes in overlap area, the 1st step of localization algorithm ends. In the 2nd step, we calculate the position of other sensor nodes using initial RNs and RNs that position are determined in 1st step. In this step, localized area is extended to surrounded area and this algorithm iterates by considering for all nodes[59]. Accumulated error average: In this subsection, to evaluate the accuracy of the localization algorithm, the variable which is called as the accumulated error average is proposed. The accumulated error average |ε| is defined as the accumulated average of the difference between the actual and estimated position of the dumb nodes. The |ε| can be achieved by: j

ε =

∑ ( xi − x* i ) + ( yi − y*i ) 2

Fig. 12. Random distribution of unknown node

Fig. 13. Sensor localization result (known 793)

B. Results and Analysis Fig. 13 shows the result from proposed sensor localization algorithm. Sensor communication range R is set to 50 and 800 sensor nodes. In this simulation, the

2

i =1

j

(3)



975


and the number of sensors. Through this study, the localization performance for given situations were estimated. The ultimate objective of this paper was to encourage research efforts to lay down fundamental basis for the development of new advanced communication techniques for efficient underwater communication and networking for enhanced ocean monitoring and exploration applications.

positions of 793 sensor nodes are determined. Therefore the more sensors are deployed, the better sensor localization performance is obtained even if the sensor range is small.

References [1]

[2] [3]

[4]

Fig. 14. Accumulated error average

[5]

Fig. 14 shows the result of accumulated error average. From the figure, the more communication is executed, the smaller the accumulated error average is obtained. Therefore, the number of the communication is an important factor to improve the accuracy of the sensor localization algorithm. The accumulated error average is almost converged after 40 times communication; this result means that after 40 times of communications, the sensor localization algorithm becomes effective.

[6]

[7]

[8] [9]

[10]

VI.

Conclusion

[11]

In this paper, the state of the art of Underwater Acoustic Sensor Network(UASN) researches were reviewed and the challenges posed by the peculiarities of the underwater channel with particular reference to monitoring applications for the ocean environment were described. The UASNs’ main architectures were investigated and a novel multipath transmission scheme, for time-critical applications in shallow water was proposed. The proposed approach combines the powercontrol strategies with multipath routing protocols and packet recovery at the destination. Without retransmission at the intermediate nodes, pattern can achieve low end-to-end packet delay. For time-critical applications in energy constrained underwater sensor networks, it is a promising transmission scheme for a good balance between packet delay and energy efficiency. After that, the general localization algorithm in shallow water environments which is based on the terrestrial techniques was discussed. From the simulations, the localization performance can be analyzed for several different conditions. We showed that localization performance depended on four kinds of factors which are initial reference, sensor node's communication range, initial reference node's position

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]


V. Hadidi, R. Javidan, A.Hadidi, Designing an Underwater Wireless Sensor Network for Ship Traffic Control, 9th HMSV Symposium, May, 2011, Naples, Italy. M. Tubaishat, S.Madria, Sensor networks: an overview, IEEE POTENTIALS, April/may 2003, pp.20-23. Chee-yee Chong, S. P. Kumar, Sensor Networks: Evolution, Opportunities, and Challenges, Proceeding of ihe IEEE, vo1.91n.8, August 2003, pp.1247-1256. M. Stojanovic, Acoustic (underwater) communications, In J. G. Proakis, editor, Encyclopedia of Telecommunications. John Wiley and Sons, 2003. J. G. Proakis, E. M. Sozer, J. A. Rice, M. Stojanovic, Shallow wateracoustic networks, IEEE Communications Magazine, November 2001, pp.114-119. Manjula.R.B, Sunilkumar S. Manvi, Issues in Underwater Acoustic Sensor Networks, International Journal of Computer and Electrical Engineering,vol.3n.1, 2011, pp. 101-110. M. Stojanovic, Acoustic (underwater) communications, in Encyclope-dia of Telecommunications, J. G. Proakis, Ed. John Wiley and Sons, 2003. Shakkottai. S,Srikant. R,Shroff. N,Unreliable Sensor Grids,IEEE INFOCOM, 2003, San Francisco, pp. 1073–1083. Hsin. C, Liu. M,Network Coverage using Low Duty Cycled Sensors. Random and Coordinated Sleep Algorithms,IEEE/ACM, 2004,California, pp. 433–442. Akyildiz.I. F, Pompili. D, Melodia.T,Underwater Acoustic Sensor Networks. Research Challenges, Ad Hoc Networks (Elsevier), 2005, pp. 257–279. Zou. Y,Chakrabarty. K,Sensor Deployment and Target Localization Based on Virtual Forces. IEEE INFOCOM, 2003, San Francisco,pp. 1293–1303. Ravelomanana. V,Extremal Properties of Threedimensional Sensor Networks with Applications,IEEE Transactions on Mobile Computing, 2004,pp. 246–257. D. Green, Underwater Acoustic Communication and ModemBased Navigation Aids, Lecture Notes in Comp. Sci., vol. 4809, 2007, pp. 474–81. M. Erol, L. Vieira, M. Gerla, AUV-Aided Localization for Underwater Sensor Networks, Proc. Int’l. Conf. Wireless Algorithms, Sys., Apps., 2007, Chicago, IL, pp. 44–54. Z. Zhou, J. Cui, S. Zhou, Efficient Localization for Large-Scale Underwater Sensor Networks, Ad Hoc Net., vol. 8 no. 3, May 2010, pp. 267–79. J. H. Cui, Z. Zhou, A. Bagtzoglou, Scalable Localization with Mobility Prediction for Underwater Sensor Networks, Proc. 2nd ACM Wksp. Underwater Net., 2007, Montreal, Canada, pp. 2198– 2206. A. C. .Bagtzoglou, Chaotic behavior and pollution dispersion characteristics in engineered tidal embayments, Anumericalinvestigation. Journal of the American Water Resources Association, 2007, pp. 207–219. A. Novikov, A. C. Bagtzoglou, Hydrodynamic model of the lowerhudson river estuarine system and its application for water qualitymanagement, Water Resource Management, 2006, pp. 257–276. Ian F. Akyildiz, Dario Pompili, TommasoMelodia, Underwater acoustic sensor networks: research challenges, Ad Hoc Networks,vol. 3, n. 3, May 2005, pp. 257-279.


976


[20] NadjibAitsaadi, NadjibAchirt, KhaledBoussettat, Guy Pujolle, ATabu Search Approach for Differentiated Sensor Network Deployment, in Consumer Communications and Networking Conference, January 2008. [21] Dario Pompili, TommasoMelodia, An Architecture for Ocean Bottom UnderWater Acoustic Sensor Networks (UWASN), in Proc. of Mediterranean Ad Hoc Networking Workshop (MedHoc-Net), June 2004, Bodrum, Turkey. [22] Wenli. Lin, Deshi. Li, Ying. Tan, Jian. Chen, Tao. Sun,Intelligent Networks and Intelligent Systems, 2008. ICINIS '08. First International Conference on, November 2008, pp. 155 – 159. [23] Ian F. Akyildiz, Dario Pompili, TommasoMelodia, State of the Art in Protocol Research for Underwater Acoustic Sensor Networks, Proceeding WUWNet '06 Proceedings of the 1st ACM international workshop on Underwater networks, 2006,NewYork, USA. [24] E. Cayirci, H. Tezcan, Y. Dogan, V. Coskun, Wireless sensor networks for underwater surveillance systems, Ad Hoc Networks, 2004. [25] ErdalCayirci, HakanTezcan, YasarDogan, VedatCoskun, Wireless sensor networks for underwater surveillance systems, Ad Hoc Networks, vol. 4 n. 4, 2006, pp. 431–446. [26] John Heidemann, Wei Ye, Jack Wills, Affan Syed, Yuan Li, Research Challenges and Applications for Underwater Sensor Networking, in IEEE Wireless Communications and Networking Conference, April 2006. [27] Eugenio Magistretti, Jiejun Kong, Uichin Lee, Mario Gerla, A mobile Delay-tolerant Approach to Long-term Energy-efficient Underwater Sensor Networking, in IEEE Wireless Communications and Networking Conference, March 2007. [28] Jun-Hong Cui, Jiejun Kong, Mario Gerla, Shengli Zhou, Challenges: Building Scalable Mobile Underwater Wireless Sensor Networks for Aquatic Applications, IEEE Network, Special Issue on Wireless Sensor Networking, vol. 20 n. 3, May/June 2006, pp. 12-18. [29] Shuo Wang, Min Tan, Research on Architecture for Reconfigurable Underwater Sensor Networks, in Proceedings of the IEEE International Conference on Networking, Sensing and Control, March 2005, pp. 831-834. [30] Jincheng Wang, Deshi Li, Mi Zhou, DipakGhosal, Data Collection with Multiple Mobile Actors in Underwater Sensor Networks, IEEE Workshop on Delay/DisruptionTolerant Mobile Networks (DTMN), June 2008, Beijing. [31] Saleh Ibrahim, Jun-Hong Cui, RedaAmmar, Surface-Level Gateway Deployment For Underwater Sensor Networks, Proceedings of IEEE Military Communications Conference, October 2007, Orlando, Atlanta, USA. [32] Antonio Caruso, Francesco Paparella, Luiz F. M. Vieira, MelikeErol, Mario Gerla, The Meandering Current Mobility Model and its Impact on Underwater Mobile Sensor Networks, in the 2008 IEEE INFOCOM conference, April 2008, Phoenix, AZ. [33] Jules Jaffe, Curt Schurgers, Sensor Networks of Freely Drifting Autonomous Underwater Explorers, WUWNet’06,September 2006, Los Angeles,USA, pp. 93– 96. [34] Kulkarni. I.S, Pompili.D, Task allocation for networked autonomous underwater vehicles in critical missions Selected Areas in Communications, IEEE Journal, vol.28n.5, June 2010, pp.716-727. [35] W. Cheng et al., Underwater Localization in Sparse 3D Acoustic Sensor Networks, Proc. IEEE INFOCOM, April 2008, Phoenix, AZ. [36] I. Vasilescu et al., Data Collection, Storage, and Retrieval with an Underwater Sensor Network, ACM Conf. Embedded Net. Sensor Sys., November 2005, San Diego, CA. [37] J. Jornet, M. Stojanovic, M. Zorzi, Focused Beam Routing Protocol for Underwater Acoustic Networks, Proc. IEEE INFOCOM, April 2008, Phoenix, AZ. [38] D. Pompili, T. Melodia, I. F. Akyildiz, Routing Algorithms for Delay-Insensitive and Delay-Sensitive Applications in Underwater Sensor Networks, Proc. ACM MobiCom, September 2006, Los Angeles, CA. [39] W. Zhang, U. Mitra, A Delay-Reliability Analysis for Multihop Underwater Acoustic Communication, Proc. ACM Int’l. Wksp. Underwater Net., September 2007, Montreal, Canada.

[40] W. Zhang, M. Stojanovic, U. Mitra, Analysis of a Simple Multihop Underwater Acoustic Network, Proc. ACM Int’l. Wksp. UnderWaterNet., September 2008, San Francisco, CA. [41] S. Ibrahim, J.-H. Cui, R. Ammar, Surface-level gateway deployment for underwater sensor networks, in Proc. IEEE Military Communication. Conference, October 2007, pp. 1–7. [42] W. K. Seah, H.-X. Tan, Multipath virtual sink architecture for underwater sensor networks, in Proc. OCEANS 2006 Asia Pacific Conference, May 2006, Singapore, pp. 1–6. [43] Z. Zhou, J.-H. Cui, S. Zhou, Localization for large scale underwater sensor networks, in Proc. IFIP Netw. , May 2007, pp. 108–119. [44] S. K. Das, A. Mukherjee, S. Bandyopadhyay, D .Saha, K. Paul, An adaptive framework for QoS routing through multiple paths in ad hoc wireless networks, J. Pa rallelDistrib. Comput. , vol. 63 no. 2, February 2003, pp. 141–153. [45] D. Ganesan, R. Govindan, S. Shenker, D. Estrin, Highly-resilient, energy efficient multipath routing in wireless sensor networks, AC M SIGMOBILE Mobile Comput. Commun. Rev., vol. 5 no. 4, 2001, pp. 11–25. [46] A. Tsirigos, Z. J. Haas, Multipath routing in the presence of frequent topological changes, IEEE Commun. Mag. ,vol. 39 no. 11, November 2001, pp. 132–138. [47] Lee .K. H, Yu .C. H, Choi .J. W,Seo.Y. B,ToA Based Sensor Localization In Underwater Wireless Sensor Networks,47thSICE Conference (SICE’08), 2008,Tokyo, pp. 1357-1361. [48] K. H. Lee, C. H. Yu, J. W. Choi, Y. B. Seo, ToA Based Sensor Localization Algorithm In Underwater Wireless Sensor Networks, Journal of Institute of Control, Robotics and Systems , vol.. 15 n.6, June 2009, pp. 641-648. [49] A. Mellouk, S. Ziane, P. Lorenz, A Swarm Quality of Service Based Multi-Path Routing Algorithm (SAMRA) for Wireless Ad Hoc Networks, International Review on Computers and Software (IRECOS), Vol. 1. n. 1, July 2006, pp. 11 – 19. [50] N. Reyes, J. Mahnke, I. Miloucheva, K. Jonas, Multicast Retransmission Strategies for Content Delivery in Heterogeneous Mobile Internet Environments, International Review on Computers and Software (IRECOS), Vol. 1. n. 2, September 2006, pp. 137 – 145. [51] L. Kaddour-El Boudadi, J. Vareille, P. Le Parc, N. Berrached, Remote Control on Internet, Long Distance Experiment of Remote Practice Works, Measurements and Results, International Review on Computers and Software (IRECOS), Vol. 2. n. 3, May 2007, pp. 208 – 216. [52] F. N. Sibai, On Cache Sizing in Multi-Core Cluster Architectures, International Review on Computers and Software (IRECOS), Vol. 2. n. 3, May 2007, pp. 235 – 241. [53] A. Sameh, A. Mahmoud, S. El-Kassas,RARAN: Authenticated Routing for Ad Hoc Networks Protocol with Reputation, International Review on Computers and Software (IRECOS), Vol. 2. n. 5, September 2007, pp. 463 – 474. [54] F. R. Armaghani, S. Khatun, S. S. Jamuar, M. F. A. Rasid, TCP Throughput Optimization Over 802.11 MAC Protocol in Multihop Ad-hoc Networks,International Review on Computers and Software (IRECOS), Vol. 3. n. 4, July 2008, pp. 356 – 362. [55] N. Enneya, M. El Koutbi, A New Mobility Metric for Evaluating Ad Hoc Network Performance, International Review on Computers and Software (IRECOS), Vol. 3. n. 5, September 2008, pp. 506 – 514. [56] S. M. Mazinani, M. H. Yaghmaee, F. Tashtarian, M. T. Honary, J. Chitizadeh, On Global Clustering Algorithm: Layer-Oriented Approach for First/Last Node Dying Applications in Wireless Sensor Networks, International Review on Computers and Software (IRECOS), Vol. 4. n. 2, March 2009, pp. 229 – 240. [57] S. M. Mazinani, M. H. Yaghmaee, M. T. Honary1,2, F. Tashtarian4, J. Chitizadeh, A New Middleware for Multipurpose Applications in Wireless Sensor Networks, International Review on Computers and Software (IRECOS), Vol. 4. n. 2, March 2009, pp. 241 – 253. [58] KecharBouabdellah, SekhriLarbi, Rahmouni Mustapha Kame, A Cross-Layer Design Approach to Achieve Energy Saving and Low Latency in MAC Layer for Delay Sensitive Wireless Sensor Network Applications, International Review on Computers and Software (IRECOS), Vol. 4. n. 3, May 2009, pp. 308 – 320.



977


[59] F. Tashtarian, A. T. Haghighat, M. H. Yaghmaee, S. M. Mazinani, M. T. Honary, On Global Clustering Algorithm: LayerOriented Approach for Multi Hop Wireless Sensor Networks, International Review on Computers and Software (IRECOS), Vol. 4. n. 5, September 2009, pp. 567-576. [60] NesaMouzehkesh, Nor K. Noordin, MohdFadlee A. Rasid, Proactive Traffic Adaptive Tuning of Contention Window for Wireless Sensor Network Medium Access Control Protocol, International Review on Computers and Software (IRECOS), Vol. 5. n. 1, January 2010, pp. 6-13.

Manijeh Keshtgary is the head of Dept. of Computer Eng. & IT, Shiraz University of Technology, Shiraz, Iran. She received her Master’s degree in Electrical & Computer Eng. from Colorado State University, CSU, Fort Collins, USA in 1993 and her PhD degree in Computer Eng. from Sharif University of technology in 2005. Dr.Keshtgary’s research interests include MANET, Wireless Sensor Networks and GSM security issues. Amin Hadidi was born in 1984. He is graduated from MSc Degree in Information Technology Computer Engineering (Machine Intelligence and Robotics) from Amirkabir University of Technology, Tehran, Iran, in 2011. His major fields are sensor networks, ad-hoc networks and sonar systems.

Authors’ information 1 Department of Computer Engineering, Islamic AzadUniversity, Dezfoul Branch. E-mail: [email protected] 2 Department of Computer Engineering and Information Technology, ShirazUniversity of Technology. E-mail: [email protected] 3

The head of Department of Computer Engineering and Information Technology, Shiraz University of Technology. E-mail: [email protected] 4 Fars Regional Electrical Company. E-mail: [email protected]

Vahid Hadidi was born in 1980. He is a student of MSc Degree in Computer Engineering (Hardware) at Islamic Azad University DezfoulBranch, Iran.His major fields are sensor networks, ad-hoc networks and sonar systems.

Reza Javidan was born in 1970. He is graduated from MSc Degree in Computer Engineering (Machine Intelligence and Robotics) from ShirazUniversity in 1996. He received Ph.D. degree in Computer Engineering (Artificial Intelligence) from Shiraz University in 2007. His major fields are artificial intelligence, image processing and sonar systems. Dr. Javidan is now assistant professor and lecturer in Department of Computer Engineering and Information Technology in Shiraz University of Technology.



978


Creating Barcode Using Fractal Theory K. Kiani, E. Kosari, M. R. Simard Abstract – This article introduces and explains a new method of 2D coding based on L-Systems Fractal Hypothesis which has a better security and capacity for coding. The coding, image processing and decoding of two dimensional codes by Hopfield Neural network or a simple decision function are studied step by step in this article. Different case studies for a new barcode shape show that the current approach provides more reliable method in the early stage of design process. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Barcode, Fractal, Image Processing, L-System

I.

[5]. Since that publication, many researchers have worked on this problem [6], among them Hilbert [7] found some of the simplest curves in the 2-D discrete space, called the Hilbert curves. These curves have been used in quite a lot of applications so far, that include: representation of 2-D patterns [8], mapping a multidimensional space into 1-D space for output display [9], scanning for pseudo-color image display [10], texture analysis [11], data compression [12], and so on [13],[18]. L- System is presented by Lindenmayer, a biologist, in 1968, to model the growth processes of Plant development and morphology of a variety of Multi cellular organisms. It can also be used to generate selfsimilar fractals such as iterated function systems [14] In this article we will show the Fractal Coding procedure by L-system. The advantage of using Fractal curves is that its’ shapes could be used in defining 2D Codes in a favorable security and capacity. Also the Hopfield Neural Network is used for decoding of Fractal Codes by recognizing the lines. A Hopfield network is an integrated artificial neural network that serves as contentaddressable memory systems with binary threshold units. By any input the most similar stored pattern in network will be red and chosen as output. Its function is very similar to human brain in remembering an image or a memory by watching a part of it. Other method is to check if there is a line exists on the fractal curve or not by using a simple decision function based on checking a few points along with the line and its neighbors.

Introduction

A barcode is an optical machine-readable representation of data, which shows data about the object to which it attaches. By prevailing of barcodes for their good accuracy and quickness, storing more data became easier than before. Many studies have accomplished on increasing barcode data storing capacity suggesting the increase of barcode digits or defining multiple barcodes by setting them adjacently. However, mostly had problems; including the increase of coding costs, barcode size and difficulties in reading and decoding. Following these problems, the idea of 2D barcodes is raised. 2D barcodes are able to store more data in smaller sizes. There are some different types of 2D barcodes such as QR-Code, Maxi Code, Data Matrix and PDF 417. One of its more common types is PDF 417 which is not like barcodes in shape consisting Black-White meshes similar to Newspaper puzzle. It can store plenty of data in a small dimension as a portable data bank. Although, increasing data volume will increase the barcode size [1]. In 1994, a Japanese company invented a new coding system; QR. QRs are 2D matrices that were first invented in manufacturing trucks but nowadays they are utilized for other applications because of their good storing capacity and 7-30% capability of data recovery with no problem and confusion by any damage [2].This technique stores plenty of data in an image. By scanning the image and using decode process the data will be extracted. Data matrix can store a number of information in an image as a sign or label on the goods [3]. Our proposed method is using L-Systems in defining Fractal Codes. This technique provides an eligible security and complication in preventing deception while give us the capability of storing large volume of information including personal data and contact details or even a dairy of driver in codes. Fractal is an inharmonious Geometric shape which can be divided in sectors as subsets of a reference [4]. In 1890, the Italian mathematician Peano presented a family of curves that pass through all points in a space

II.

Drawing the Hilbert Curve Using L-System

A Hilbert Curve is a continuous fractal space-filling curve first described by the German mathematician David Hilbert in 1891, as a variant of the space-filling curves discovered by Peano in 1890 [4]. Below (Fig. 1) is the 6 order of Hilbert curve.


979


K. Kiani, E. Kosari, M. R. Simard

Firstly for preparing the scanned image we should preprocess it. Preprocessing: At this stage for recovering the image quality a median filter is used to remove the salt and pepper noises. A median filter is a nonSlinear digital filter which we use it by defining a 3×3 matrix widow at image. By scanning a mask for the image, each 8 pixels adjacent to the center of the window will be included in the median operator [15]. The next stage is changing recovered image to a black-white picture. Then by recognizing the lines, 0 &1 bits are determined. The scanned image will has a different size and will be angular comparing with its original image. For solving this problem fist corners of image will be found and then the turning angel will be determined and then induced reversely to the image [16]. To sizing the image, diameter of scanned image will be determined and compared with original image.

Fig. 1. A Six order Hilbert curve

III. Data Encoding Step First separate the numerical coordinates in each matrix then regarding the codes define the line between spots. Here we use fractal pattern which gives us 64 data bits. As a result the digit 1 is the sign of having the Line and digit 0 shows absence of the Line. Whereas the main codes will be defined in barcode, we need 8 bits code. For further security in coding other ways could be used for showing bit0 or 1 on the Hilbert curve. Below shows an encoded 5 order Hilbert curve.

Fig. 2. Encoded 5order Hilbert curve

Remember that 5 order of Hilbert curve has 1024 numerical coordinates which can show 170 characters. 6 order of Hilbert Curve has 4096 numerical coordinates which can show 682 characters and also there are 16834 numerical coordinates and ability of having 2730 character for 7 order Hilbert Curve.

IV.

Fig. 3. Scanned barcode with determined four corners

In above image a scanned barcode with determined four corners is shown. The amount of rotation angle is calculated with following equations: 1 2

Barcode Reader

(1)

1, 2

It’s an instrument that transfers the image of barcode to a computer for processing and decoding. There are three models barcode readers. Fixed barcode reader, portable barcode reader and wireless barcode reader. We have used a 2D fixed scanner which scans the image of barcode for processing in a computer.

V.

/ /

Then black points of the image then morphologically with operator of Erosion, a basic operator, defined in a 5×5 matrix window and implicated to image. This will strengthen black lines of image to be better determined [15]. The decoding process is accomplished using two different methods: In the first method the redundant margins of the scanned 2d barcode image is removed and then some simple preprocessings apply to stimulate the lines and remove any noise.

Decoding Barcode

Decoding is reading of data from scanned image as bits and then excluding the primary raw information. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


980


using a Hopfield neural network and a simple decision function has been explained. Considerable amount of data can be stored in various orders of Hilbert fractal barcode; however, size limitation should be considered The application of the approach is exemplified by numerical solutions provided to support the analysis. Results show significant capability in implementing the proposed method. Fig. 4. Preprocessed barcode image

References

Then using the specified Hilbert fractal function we match the prepared image to its Hilbert fractal sketch. Finally every bit code is extracted using a simple decision function which decides whether there is a line exists or not. This decision is taking place by reading a few sample points along the supposed line location [17]. The second method of decoding just differs in the decision function. Here we use a Hopfield neural network for this purpose. Hopfield neural network will be applied for storing some patterns and their recovering. Training method of this neural network will be as follow: first 2D images of the line patterns consists of vertical and horizontal lines with different kind of noises will be introduced into the neural network. Resulted vectors will give us the weighting function of neural network as below equation: the main advantage of using this neural network is to make decoding process more reliable against noises and mismatches between scanned image and fractal pattern image: 1

[1] [2]

[3]

[4]

[5] [6] [7] [8]

[9]

(2) [10]

Calculating the weighting vectors which is a 121×121 matrix (as same as scanned image), 11×11 sub-images deduces from scanned image will be applied in neural network. The neural network will map the input image to one of its previously stored images and so the input image is classified [17]. Then the process output is a bit stream which the data codes can be extracted after converting every 8 bits into its Ascii character and so the decoding is completed. We have had a comparison between two methods while introducing more noise to the scanned image as well as other capturing uncertainties. These will chaos the barcode lines width in one image and introduce some shrinkages and gaps in lines. Obviously the extracted bit codes using Hopfield neural network has much lower errors than using the simple decision function method.

[11]

[12]

[13]

[14]

[15] [16]

[17]

VI.

Conclusion

In this paper we proposed a new barcode using fractal L systems. Also the implementation process including barcodes encoding, barcodes reading, enhancement of scanned image and decoding Hilbert fractal barcodes

[18]


PEI Jiao, LI Fang-wei, Two-dimensional Bar Code PDF417 Decoding Technology , Journal of Chongqing, 2003. Ohbuchi, E.; Hanaizumi, H.; Hock, L.A., Barcode readers using the camera device in mobile phones, Cyberworlds, 2004 International Conference on, vol., no., pp. 260- 265, 18-20 Nov. 2004. Ouaviani, E.; Pavan, A.; Bottazzi, M.; Brunelli, E.; Caselli, F.; Guerrero, M., A common image processing framework for 2D barcode reading, Image Processing and Its Applications, 1999. Seventh International Conference on (Conf. Publ. No. 465) , vol.2, no., pp.652-655 vol.2, 1999. Barnsley‚ M. F.‚ Demko‚ S.‚ Iterated function systems and the global construction of fractals‚ The Proceedings of the Royal Society of London A399 (1985) 243–275. S. J. Peano, “Sur une courbe qui remplit touteune aire plane,” Math.Annu., vol. 36, 1890. S. J. Butz, “Alternative algorithm for Hilbert’s space filling curve,” IEEE Trans. Comput., vol. C-20, pp. 424–426, Apr. 1971. S. J. Gardner, “Mathematical games,” Scientif. Amer., pp. 124– 133, Dec.1976. S. J. Abend and T. J. Harley, “Classification of binary random pattern,”IEEE Trans. Inform. Theory, vol. IT-11, pp. 538–544, Oct. 1965. T. J. Patrick, D. R. Anderson, and F. K. Bechtel, “Mapping multidimensional space at one dimensional space for computer output display,” IEEE Trans. Comput., vol. C-17, pp. 949–953, Oct. 1968. F. K. Stevens, A. F. Lehar, and F. H. Preston, “Manipulation and presentation of multidimensional image data using the Peano scan,” IEEE Trans. Pattern Anal. Machine Intell., vol. PAMI-5, pp. 520–526, Sept. 1983. F. H. Sonlinh, T. K. Oh, N. Caviris, Y. Li, and H. H. Szu, “Texture analysis by space-filling curves and one dimensional Haar wavelets,” Opt.Eng., vol. 31, no. 9, pp. 1899–1906, Sept. 1992. H. Ansari and A. Fineberg, “Image data ordering and compression using Peano scan and LOT,” IEEE Trans. Consumer Electron., vol. 38, pp. 436–445, Aug. 1992. Maria Petrou, Peixin Hou, Sei-ichiro Kamata, and Craig Ian Underwood, “Region-Based Image Coding with Multiple Algorithms”, IEEE Trans.on Geoscience and Remote Sensing, vol.39, No3,Mar 2001. P Prusinkiewicz, Graphical applications of L-systems, In Proceedings on Graphics Interface '86/Vision Interface '86, Canadian Information Processing Society, Toronto, Ont., Canada, Canada, 247-253. R. Gonzales, R.Woods, Digital Image Processing, (Prentice Hall), second edition, 2001. M. Unser, P. Thevenaz, L. Yaroslavsky, Convolution based interpolation for fast, high-quality rotation of images, IEEE Transactions on Image Processing vol.4, no.10, 1995, pp.13711381. Ovidiu Pârvu, Andrei G. Bălan , A method for fast detection and decoding of specific 2d barcodes. 17th Telecommunications forum TELFOR 2009. Zahra Sadri Tabatabaie, Rahmita Wirza Rahmat, Nur Izura Udzir, Esmaeil Kheirkhah, Adaptive Skin Color Classification Technique for Color-Based Face Detection Systems using Integral Image, International Review on Computers and Software (IRECOS), Vol. 6 N. 1, 2011, pp. 32-39.


981


Authors’ information Kourosh Kiani received the BS and M.Sc. Degrees in electrical engineering from Delft University of Technology in Delft, the Netherlands in 1993 and the PhD degree in Medical Information from Erasmus University in Rotterdam, the Netherlands in 1997. Between 1997 and 2005 he was employed as an engineer and researcher by Philips Medical Systems. He left industry in 2005 to become Assistant Professor in the Department of Electrical Engineering at the Semnan University of Semnan, in Iran and in the Department of Electrical Engineering at the Amirkabir University of Tehran, in Iran. Ehsan Kosari resived the BS Degree in control engineering from Imam Khomeini International University in Qazvin in 2008 and he is a MSC student in the field of electronic engineering in Semnan University in Iran . He has one paper about barcode generation using fractal theory.His reaserch interest includes : Image processing,Pattern Recognition,Chaos Theory . Mohammad Reza Simard resived the BS Degree in control engineering from Imam Khomeini International University in 2008 and he is MSC student in the field of communications in Azad University in Iran . He has one paper about barcod generating using fractal theory.His reaserch interest includes: Image Proccessing, Pattern Recognition, Chaos Theory .



982


Model Proving of Urban Traffic Control Using Neuro Petri Nets and Fuzzy Logic Rishi Asthana1, Nilu Jyothi Ahuja2, Manuj Darbari3

Abstract – The development of control systems to handle the congestion at intersection in urban traffic is a critical research issue. Petri Nets and Fuzzy Logic have played vital role in the development of such systems. Petri nets, a mathematical modeling tool, makes graphical modeling, simulation and real time control modeling, more functional. Fuzzy logic deals with uncertainties in the environment to make the systems more realistic. In this paper, a real time traffic control model has proposed and implemented in MATLAB. Several results have been discussed and found satisfactory. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Fuzzy Logic, Petri Nets, Neuro Petri Nets, Membership Functions (MFs)

I.

The PN model ensures the following advantages in traffic control system modeling 1.Modularity: The whole model is made up of individual sub models. The supports easy model development and readability improvement. 2. Functionality: The model developed with Petri Nets is suitable to be used in simulation environment and become more functional. This paper is divided into 4 sections. In section II, the basic concepts of traffic control systems are discussed. In section III, the neural and fuzzy Petri nets are discussed. Section IV gives the outlook of the proposed system. Section V presents the implementation of the proposed systems with MATLAB.

Introduction

The development of control systems to handle the congestion at intersection in urban traffic is a critical research issue. The major requirements of the developed system are, the signal must not allow the ambiguous movement to the traffic and it must be clear that how/when the indication of signal shown to be changed. Two other aspects to be handled are to take decisions about signal indication sequence in the control system to make the system well optimized and development of control logic for signal generation. Several mathematical algorithms, control strategies, programming are integrated with soft computing techniques, like Fuzzy Logic, Neural Network, Petri Nets etc. for the development of automatic traffic control systems. SCOOT [1] is the major real time traffic control systems. Several systems have been developed for traffic control systems, like SCATS [2], PRODYN [3], OPAC [4], UTOPIA [5], TUC [6] etc. The problem of traffic control aims to be implemented as simulation and control logic. The logic behind control strategy is explain by several techniques like pseudo code and flow chart that is represented in the form of objective function or as set of constraints and specified relevant rules Petri nets [7],[8] is the technique that gives clear means for presenting simulation and control logic and also supports the generation of control code from the PN graph. The Petri net based models supports the direct generation of control code with help of graphical tools such as NETMAN [9] and UltraSAN [10]. Also, the use of Petri Nets (PN) ensures the control model to enforce all the safety rules. It also ensures the acceptability of all the signal plans. Colored Petri Net (CPN) are used to model traffic control system as a scheduler [11],[12],[13].

II.

Control of Traffic Signal

Movement and phase are the two terms that are very important in traffic signal control theory. A movement is defined as a specific traffic flow that occurs at the intersection. Phases are the paired combinations of the movements’. For example, when a phase is selected greens are displayed for both of the movements involved, while the other will receive the RED signal. The signal combination or phases are found like, (T1,T3), (T2,T4), (R1,R3), (R2,R4). At a time one phase will get the green signal and other will receive the red signal. The optimum design methodology of the signal control system is that no vehicle should stop at the Intersection. Several Kinds of control strategies have been identified: 1. Determined: In this control strategy, the phase sequence are fixed and durations are also fixed. 2. Fully actuated: The phase sequence depends on the demand of traffic, detected by the sensor technology.


983


R. Asthana, N. J. Ahuja, M. Darbari

3. Semi Actuated: The mirror street phase are demand responsive. 4. Queue management: Queue length will decide the signal switching to be happened. Number of parameters is required to be identified for the implementation of actual traffic signal control system: 1. Yellow time (YT) minimum duration that a phase has to remain green before any switching can be made. 2. Green Time (GT): Maximum duration that a signal can display green. 3. E-Interval (E-I) Time: The time that a green is extended for each vehicle arrived at the detector from the instant of the arrival at the detector.

Here: P={p1, p2, …………..,pn} Finite set of places T={t1, t2, t3, …………, tn} Finite set of Transitions D={d1, d2, ……………, dn} Finite set of propositions P∩T∩D=Φ |P|=|D| I and O were the functions of set of input and output places of transitions, where I: P->T was the input function, a mapping from transmission to bags of places. The modeling approach by Fuzzy Petri Net is discussed in Fig. 2.

Fig. 2. Fuzzy Modeling through Petri Nets

The neural extension of the Petri Nets has three components, places, transitions and directed arcs [18]. The directed arcs are connecting the places to the transitions and transitions to the places. There are no arcs connecting transitions to transitions or places to places. A simple graph of the Neural Petri Net is given below.

Fig. 1. X-Junction showing places and movements

III. Neural Fuzzy Petri Nets and Modeling Petri Nets [15] are also known as a place/Transition Net or P/T net. This is the mathematical modeling languages for the description of Discrete Event Systems (DES) PN theory is developed 1962 by Carie Asam Petri. They are highly applicable ion graphical modeling, Mathematical modeling, simulation and real time control by the use of places and transitions. But most of the real world systems are requiring the modeling systems that are capable to handle uncertainty in systems properly. To deal with this issue a fuzzy extension of PNs is proposed called Fuzzy Petri Nets (FPN) because fuzzy logic is an excellent theory capable to handle uncertainties. Several types of fuzzy Petri nets are identified: 1. Basic Fuzzy Petri Nets (BFPN), 2. Fuzzy Timed Petri Nets (FTPN), 3. Fuzzy Colored Petri Nets (FCPN), 4. Adaptive Fuzzy Petry Nets (AFPN), Composite Fuzzy Petry Nets (CFPN). As discussed in [16], [17], the fuzzy petri nets are defined by 8 tuples FPN= (P, T, D, I, O, f, α, β ):

Fig. 3. Neural extension of Petri Nets

The above architecture includes, input layer composed of p inputs, hidden layer composed of hidden transitions, output layer composed of q output places. In this model, the transitions are acting like processing units. At the output layer, a plan of traffic control to decide



984


the phases and movement at an Intersection. At the input layer, the fuzzy variables are inputted for processing to get the desired outcome in terms of fuzzy output variables.

IV.

V.

MATLAB Simulation

For MATLAB SIMULATIONof the proposed traffic control systems, we have developed three input membership functions and two output fuzzy membership functions. At input, we have Traffic Flow (TF), Traffic Queue (TQ) and Incoming Flow (IF) and at output the membership functions are, Green Time (GT) and Extended Interval Time (E-IT).

Proposed Fuzzy Logic Control Systems

To extend the capability of TCS of the traffic management at intersection, fuzzy techniques are integrated. Fuzzy Traffic Control Systems (FTCS) uses sensors that not only indicate the presence of vehicle, but also sensors gives estimation of traffic densities in the queue. The placement of sensors at traffic intersection in Fig. 1 can be explained as follows.

Membership Function Definitions The definition of these membership functions are defined below.

VL-Very Low, L-Low, N-Normal, H-High, VH-Very High Fig. 6. MF of Traffic Flow (TF) Variable

Fig. 4. Sensor placement at X Junction

Three types of electromagnetic sensors are put on the intersection. 1. Front Sensor, 2. Middle Sensor, 3. Rear Sensor. Front Sensor counts the number of vehicles passes the traffic light. Middle sensor counts the traffic density in the di from the signal. Rear sensor counts the incoming number of vehicles towards the intersection. Now the fuzzy logic controller is responsible for controlling the length of green time. A traffic control system model has been proposed in [19]. It is the Neural Fuzzy Petri Net [20], [21] model implementation of traffic control systems.

L-Low, H-High, VH- Very High, EH- Extremely High Fig. 7. MF of Incoming Flow (IF)

S-Small, L-Large, EL- Extra Large, OF-Over Flow Fig. 8. MF of Traffic Queue (TQ)

L-Less, M- More, EM-Extremely More, UE- Unacceptable Fig. 5. N-Dimensional Self Organized Petri Net Modeling Urban Traffic Control System

Fig. 9. MF of the Green Time (GT)



985


L-Low, H- High, EH- Extremely High, U-Unacceptable Fig. 10. MF of the Extended –Interval Time (E-IT)

Knowledge Base definition The knowledge base of the system is defined by the following rule matrices.

Fig. 11. MATLAB Fuzzy Inference System (FIS)

TABLE I RULE MATRIX 1 Traffic Flow (TF)

Incoming Flow (IF)

Green Time (GT)

VL

L

L

L

L

L

N

L

M

H

H

L

L

VH

EM

VH

H

M

TABLE II RULE MATRIX 2 Time Delay (TD)

Extended Interval Time (E-IT)

VS

L

S

L

L

H

EL

EM

Fig. 12. Rule Editor in MATLAB for proposed system

VI.

In this paper, a Urban Traffic Control System has been proposed and modeled using Neural Petri Nets. Also, It is implemented by MATLAB SIMULATION. Several results have been carried out and found satisfactory. In future, our efforts would be dedicated to enhance the proposed system by extending its capability in terms of optimization of traffic control plans using nature inspired computing, i.e. ACO, PSO etc.

The following results have been obtained by the simulation and the results are found satisfactory.

S. No. 1 2 3 4 5 6

TABLE III SIMULATION RESULT I Input1 Input2 Output [GreenTime] [Traffic Flow] [IncomingFlow] 10 5 20 10 10 27 20 17 50 40 25 100 60 35 135 70 45 165

References [1]

[2]

TABLE IV SIMULATION RESULT II

[3]

S. No.

Input [Traffic Queue]

Output [E-I Time]

1

5

20

2

10

35

3

20

50

4

25

60

5

35

75

6

40

120

Conclusion and Future Scope

[4]

[5]

[6]


R. Robertson and R. D. Bretherton, “Optimizing networks of traffic signals in real time—the SCOOT method,” IEEE Trans. Veh. Technol., vol. 40, pp. 11–15, Feb. 1991. G. Sims and K. W. Dobinson, “The Sydney coordinated adaptive traffic (SCAT) system philosophy and benefits,” IEEE Trans. Veh. Technol., vol. VT-29, pp. 130–137, May 1980. J. J. Henry, J. L. Farges, and J. Tuffal, “The PRODYN real time traffic algorithm,” in Proc. 4th IFAC/IFORS Conf. Control Transportation Systems, Baden-Baden, Germany, 1983, pp. 305– 309. N. H. Gartner, P. J. Tarnoff, and C. M. Andrews, Evaluation of optimized policies for adaptive control strategy, in Transportation Research Record 1324, Trans. Res. Board, Washington, DC, 1991. C. DiTarnato and V. Mauro, UTOPIA, in Reprints of Control, Computers, Communications and Transportation (CCCT 89), IFAC, IFIP, IFORS, Paris, France, 1989. V. Dinopoulou, C. Diakaki, and M. Papageorgiou, “Simulation


986


[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18] [19]

[20]

[21]

investigation of the traffic-responsive urban control strategy TUC,” in IEEE Proc. Intelligent Transportation Systems Conf., Dearborn, MI, 2000, pp. 458–463. G. F. List and R. Troutbeck, “Advancing the frontier of simulation as a capacity and quality of service analysis tool,” Transport. Res. Circ., pp. 485–502, 2000. F. DiCesare, P. Kulp, M. Gile, and G. F. List, “The application of Petri nets to the modeling, analysis and control of intelligent urban traffic networks,” in Proc. 15th Int. Conf. Application Theory Petri Nets, R. Valette, Ed., Zaragoza, Spain, June 1994, pp. 2–15. J. L. Gallego, J. L. Farges, and J. J. Henry, “Design by Petri nets of an intersection signal controller,” Transport. Res., pt. C, vol. 4, no. 4, pp. 231–248, 1996. W. H. Sanders, W. D. Obal II, M. A. Qureshi, and F. K. Widjanarko, “The UltraSAN modeling environment,” Perform. Eval., vol. 24, no. 1, pp. 89–115, 1995. K. Jensen, “Colored Petri nets,” in Petri Nets: Central Models and Their Properties, ser. Lecture Notes in Computer Science . New York: Springer-Verlag, 1986, vol. 254, pp. 248–299. Cois, A. Fanni, and A. Giua, “An expert system for designing and supervising a discrete event model,” in Proc. IEEE Int. Workshop Intelligent Motion Control, Istanbul, Turkey, 1990, pp. 103–107. Giua, “A traffic light controller based on Petri nets, CML final project report,” Dept. Elect., Comput., Syst. Eng., Rensselaer Polytechnic Inst., Troy, NY, 1991. G. F. List, M. Setin, Modeling traffic signal control using petri nets, IEEE Transactions on Intelligent Transportation Systems, Vol. 5, No. 3, Sept. 2004. R. Zurawski, M. Zhou, Petri nets and industrial applications: A tutorial, IEEE Transactions on Industrial Electronics, 41(6), 567583, Dec. 1994. M.H. Aziz, E. L. J. Bohez, M. Parnichkum, C. Saha, Classification of fuzzy petri nets and their applications, World Academy of Science, Engineering & Technology, 72, 2010, 394401. Yung-Hsiang Cheng, and Li-An Yang, “A Fuzzy Petri nets approach for railway traffic control in case of abnormality: Evidence from Taiwan railway system,” Expert Systems with Applications, vol. 36, pp. 8040- 8048, 2009. Seely, Petri Net implementation of neural network elements, M. Sc. Thesis, Nova Sourthastern University, 2002. M. Darbari, R. Asthana, H. Ahmed, N. J. Ahuja, Enhancing the capability of N-Dimension Self-organizing Petri Net using NeuroGenetic Approach, IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 3, No. 1, May 2011. A. Jabeur Telmoudi, Lotfi Nabli, Radhi M’hiri, Modeling Method of Robust Control Laws for Manufacturing System to Temporal and non Temporal Constraints through Petri Nets, International Review on Computers and Software (IRECOS), Vol. 4. n. 2, pp. 266 -277, March 2009. Abhay Kumar Srivastava, Manuj Darbari, Hasan Ahmed,Rishi Asthana, Capacity Requirement Planning Using Petri Dynamics, International Review on Computers and Software (IRECOS), Vol. 5 N. 6, pp. 696-700, November 2010.

Authors’ information Rishi Asthana is currently working as an associate professor in the Dept. of Electrical Engineering at Babu Banarasi Das University, Lucknow. He is pursuing Ph. D. from University of Petroleum and Energy Studies, Dehradun, India and have a teaching experience of more than fifteen years. Prior to his current assignment, he has taught for five years in Pauri Garhwal as lecturer and ten years in B.B.D.N.I.T.M. Lucknow in different positions. He has published fifteen papers in referred international and national journals. His area of interest includes Control Systems, System Modeling and Soft Computing. Dr. Neelu Jyothi Ahuja is working as Assistant Professor in the College of Engineering Studies, University of Petroleum & Energy Studies, Dehradun. Her area of interest includes Computer Sciences- Expert Systems, Artificial Intelligence, Object Oriented Development, Programming Languages. She has 13 years experience and published many papers in journals and conferences at International and National level. Dr. Manuj Darbari is currently working as an associate professor in the Dept. of Information technology at B.B.D.N.I.T.M (Babu Banarasi Das National Institute of Technology And Management), Lucknow. He holds a Ph. D. from Birla Institute of Technology Mesra Ranchi, India and having a teaching experience of more than eleven years. Prior to his current assignment, he has taught for one year in M.N.R.E.C Allahabad as lecturer and ten years in B.B.D.N.I.T.M. Lucknow in different positions. He has published fifteen papers in referred international and national journals. He is selected for marquis who’s who in science and engineering 2003‐2007. His teaching areas are information science, ERP, software engineering, workflow management



987


A New Evaluation Strategy Based on SVM for Supply Chain Finance Wenfang Sun1,2, Jindong Zhang2, Xilong Qu3

Abstract – The concept of supply chain finance is introduced, The SAW( Self-avoiding random walk) model characterizing fractal dimension of biological macro molecular fractal network is analyzed. Support vector machine algorithm has been adopted to become an evaluation strategy for supply chain finance. The evaluation and choice for kernel function is very important. In this paper, a quantity estimate is proposed though empirical risk and confidence interval based on the structural risk theory. The results of simulation experiments are shown, and the feasibility and effectiveness of this method is proved. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Supply Chain Finance, Kernel Function, Support Vector Machine

I.

support vector machine algorithm [2][3], kernel Fisher classifier[4], kernel principal component analysis [5], kernel independent component analysis [6] tec. Among them, typical algorithm is support vector machines. When using support vector machines to solve practical problems, select the appropriate kernel is a key factor [3]. Currently there are a number of ways to the construction of kernel functions，Ling Zhang [7] proved that for a given sample kernel function must exists. But there is still no uniform guideline yet on the kernel function and parameter selection [8], select it still by experiment. Commonly using cross-validation method [8] and leave-one method [9] in a limited data set as parameter optimization. Literatures [10] made a useful discussion on selecting parameters of kernel function，but principally used cross-validation method [11] and leave-one method [12] in a limited data set for parameter optimization. In this paper, using the principle of structural risk minimization [2] and projection analysis, to make a quantitative analysis on the kernel function to decision function’s influence, to guide the kernel’s option, has some innovative. This paper focuses on Vapnik proposed standard support vector machine, but after a little changes also applies to various support vector machines.

Introduction

The reverse factoring process begins with the supplier sending an invoice to the buyer. The buyer approves the invoice and uploads it to the Supply Chain Finance platform, thereby creating an irrevocable payment obligation. The supplier is now able to sell the invoice (i.e asset based finance) to the financier at an attractive rate, based on the buyer risk. The payables financing transaction is designed to be a “true sale” where the risk is transferred from the supplier to the financier. For example, in a supply chain with 75-day terms, it is possible for the supplier to receive payment on day 5, at a low cost, while the buyer, thanks to supply chain financing, is able to pay the financier on day 75.The success of a Supply Chain Finance (payables financing) solution hinges on the real-time visibility and integrity of invoice data, to enable all participants in the structure to track invoices, advance payments and settlements. If the process were not automated, the key efficiencies of lending against eligible invoices would be lost in the costs of manual processing. The key phases of implementation of the vendor finance/supplier finance programmer are outline below [1]. Since 60s of last century, scholars began to study the problem of machine learning based on the data. And until the 90s of last century, Vapnik and his colleagues created a statistical learning theory (SLT), making data-based machine learning to become a more complete theory, and on this basis, finally created a class of effective generic machine learning algorithms - Support vector machines [1][2]. The application of kernel as a universal technology，can extended to other learning systems. Currently, the main kernel machine algorithm including

II.

Kernel Function Methods

For the two types of problems, given sample set, ( xi , yi ) ; xi ∈ R n , yi = ±1 , i = 1, 2, ,l , SVM algorithm trained the separating hyperplane (decision function) H as [3] f ( x ) = ( w ⋅ x ) + b , w is called weight vector，b is called offset, the output y = sgn ( f ( x ) ) , sgn is the sign


988


Wenfang Sun, Jindong Zhang, Xilong Qu

function. Decision function f ( x ) 's value is the sample x to the direction w in a straight line projection (Fig. 1), f ( x ) 's value form a one-dimensional F space. Set Nonlinear function Φ realize the input space X mapped to feature space Z (Hilbert space) (Fig. 2), , kernel function which Z ∈ Rm K ( xi ,x j ) = Φ ( xi ) ⋅ Φ ( x j ) , xi ,x j ∈ X ,which ( ⋅) is the

(

negative samples, l+ + l− = l ; S + , S − respectively as positive and negative sample set. According to Vapnik's statistical learning point of view, machine learning should be consistent with the principle of structural risk minimization (ERM) [2]，that means the expected risk of learning system from two parts, the risk of experience and confidence interval. Standard SVM through optimize the problem:

)

inner product. min τ ( w ) =

ω ,b,ξ

s.t.

yi

l 1 || w ||2 +C ξi 2 i =1

∑

( ( w ⋅ Φ ( xi ) ) + b ) ≥ 1 − ξi ,

i = 1,

,l

(2)

ξi ≥ 0 , i = 1, ,l To get the learning machine, || w ||2 / 2 in formula expresses the confidence interval, while the l

Fig. 1. If is the sample's projection in the direction of w

C

∑ ξi controls the risk of experience. After analysed, we i =1

believe that for the classification issue, performance of the kernel function depends on its influence to decision-making function' promote ability, that is a good kernel function can reduce the risk of expectations of the learning machine.

III. Self-Avoiding Random Walk (SAW) Model

Fig. 2. From the input space to feature space

Define the value of Φ ( xi ) as image for sample xi in the feature space, the sample set to feature space through nonlinear mapping is called as the image set [8]. In the feature space for the separating hyperplane can be obtained: f ( x ) = ( w ⋅ Φ ( x )) + b

(1)

After the nonlinear mapping sample set，its image set in the feature space may be divided into linear or nearly linear separable. The sample image set as input, high dimensional feature space as the input space, the traditional linear classification algorithm can be directly achieved most of nonlinear classification, such as Fisher linear discriminant [4]. Support vector machine is representative of the use of kernel techniques, but essentially, we think it is the linear discriminant algorithm. Commonly used kernel function is divided into inner product kernel function and translation-invariant kernel function, such as the Gauss RBF kernel and polynomial kernel and so on. With using the different kernel functions and parameters, the direction of the hyperplane w which trained by support vector machine is also different. To facilitate the description, define several variables and set here: zi as image Φ ( xi ) of sample xi in the feature space; l+ , l− respectively as the number of positive and


Biological macro molecular (mainly including protein and nucleic acid) has very sophisticated fractal structure, and it is shown as Fig. 1. Fractal network structure of biological macromolecules has a very fine and intelligent information, material and energy transmission mechanism. The well-known American expert polymer P. J. Flory has studied biological macromolecules strand conformation (the mercy way of macromolecular chains in the space) and gets the SAW model: Df ＝

d +2 3

( d is Euclidean dimension)

Although Flory has made a number of approximate calculation in the deducing process, but the result is correct. Later, P.J.Fisher used statistical mechanics method and proved its correctness and improved it, so it is also called Fisher－Flory formula. d is the required number of variables for the movement in the space. The Euclidean space determined by n independent variables, namely, d = n . In the ideal state, the fractal supply chain network system is a self-organizing collaborative dynamic system, so, to determine d according to the steps as follows: Given the motion parameter affecting fractal supply chain system in the ideal state is: International Review on Computers and Software, Vol. 6, N. 6

989


{ x1 ,x2 ,

,

}

x p ; ( p = 1, 2 ,.....)

At first, adopting the "adiabatic elimination method" presented by the founder of Synergetics to find the slow variables disposing fractal supply chain network system in an ideal state, and integrates the role played by fast variable into the whole role through the role played by the slow variables. The slow variables integrated are: [6]

{ x1 ,x2 ,

, xq

} ; (1 ≤ q < p )

Fig. 4. Positive and negative samples projection showed a normal distribution

Secondly, analyzes the correlation between q variables, then, classifies and integrates the slow relevant variables into a new slow variable. After the classification, the slow variables are:

{ x1 ,x2,

, xm

Set the all of the value of positive class projection as F+ , then F+ ∼ N ( µ+ ,σ + ) , µ + as the mean value of fi ,

σ + as the mean square deviation of fi , equally F− ∼ N ( µ − ,σ − ) . When the sample size is large, the mean value and mean square deviation can be estimate by the training samples projected value to estimate, namely:

} ; (1 ≤ m < q )

Thirdly, repeat the second step, then analyzes the relevance between m slow variables, until there were not relevant independent variables, and then, d = n . Finally, substitutes the determined d into formula (1), the fractal dimension D f of fractal network supply chain in ideal state will be get. Fractal dimension of fractal supply chain network is the critical parameters describing dynamic complexity of the supply chain, and it is very important for describing quantitatively for the relevance between the parameters of organizational structure of fractal supply chain and improving the operational efficiency of organizational structure of the supply chain. So, computing the fractal dimension has becomes the most important part of fractal supply chain network system structure. After mapping through kernel functions, fi is zi to the direction of the straight line projection of w in the feature space (Fig. 3). If the samples can be divided into linear or nearly linear separable, then each projection value should in the vicinity of the projection value of average value in theory. In particular, when each sample is higher, according to the Central Limit Theorem [13], the projection value can be approximated considered that obey the normal distribution (Fig. 4).

µ+ = f + =

1 l+

∑

j∈S +

1 µ− = f − = l−

∑

j∈S −

∑ ( f j − f+ )

1 l+

f j , σ+ =

j∈S +

∑ ( f j − f− )

1 f j , σ− = l−

2

(3) 2

j∈S −

Name the point of intersection of distribution curve of F+ and F− as Class Critical, its corresponding projection f c as Class Critical Point. Easy to prove, if image set is linear separable or nearly linear separable, then µ+ ≠ µ− , and f c certainly exists

and is located between µ− and µ+ . Due to space limitations, this paper on the assumption that µ+ > µ− , µ+ < µ− . Easy to know, in the point of f c has:

1 2πσ −

−

( f c − µ− )2 2σ −2

e

=

−

1 2πσ +

( f c − µ + )2

e

2σ +2

(4)

Calculate the value of f c by above formula. Record

P+ as the distribution of fi > f c ( i ∈ S + ) . P− as the

distribution of

fi < f c ( i ∈ S − ) (his paper on the

assumption that µ+ > µ− ), then has:

1

e 2πσ + ∫−∞

P+ = 1 −

P− =

fc

1

fc

e 2πσ − ∫−∞

−

−

( t − µ+ )2 2σ +2

(5)

( t − µ − )2 2σ −2

dt

dt

Fig. 3. Sample projection



990


when

fi < f c ( i ∈ S + ) ,classifier will determine this

linear discriminant, to further describe the distribution of f. The research on linear discriminant analysis can be traced back to Fisher published the classic paper [14] in 1936, he basic idea is to find a direction w makes zi as a straight line projection fi in the direction w separated as much as possible. In the case of not cause confusion can called fi as sample of one-dimensional F space. In order to compare different between kernel function, normalizes f, record gi = fi / µ+ − µ− , that is, the mean

sample was negative class, so P+ indicate the probability of positive samples were classified correctly. Similarly, P− indicate the probability of negative samples were classified correctly. The degree of linear separability of sample image set is equivalent to the probability of a decision-making function were classified correctly, on this basis can get the sample linearity LF ( w ) : LF ( w ) =

1 ( P+ + P− ) 2

(6)

distance of two types are 1. Set gi as sample of one-dimensional F space. Several variables defined in the one-dimensional F space:

The meaning of probability LF ( w ) is the probability of samples were classified correctly .When LF ( w ) > θ ( θ is the threshold value), we believe that decision-making function has good degree of linearly separable. P+ and P− are not the amount of two unrelated，they are directly related to f c . f c associated with w, while w depends on the image set distribution of sample by the kernel function mapped to the feature space, so the value of LF ( w ) is determined by the kernel function , in other words, LF ( w ) expressed the influence of kernel function on the degree of linearly separable of sample image set. Because of LF ( w ) has strong meaning of probability, also available. Definition 1 Non-linearity of sample N F ( w ) : N F ( w) = 1 −

1 ( P+ + P− ) 2

the degree of scatter of sample class sb2 : sb2 = ( µ+ − µ− ) = 1 2

the degree of scatter of positive sample class s+2 : s+2 =

s−2 =

∑

(10)

µ ⎛ ⎞ ∑− ⎜ g j − µ+ −−µ− ⎟ ⎠ j∈S ⎝

2

(11)

the degree of scatter of total sample class sw : sw = s+2 + s−2

(12)

f separated as much as possible, that is, the difference between two sample | µ+ − µ− | the bigger the better (normalized the average difference of two types are 1), the same time intensive as much as possible within all samples, that is, the smaller the degree of scatter between class sw is the better it is, on the basis can be obtained the degree of scatter between sample classes J F ( w ) :

(8)

where L is 0-1loss function, which reflects the trained learning machine on the misclassified level of training samples. N F ( w ) expressed the probability of training samples were classified incorrectly, when samples are large enough, N F ( w ) shows the risk of experience of learning machines.

IV.

2

(7)

l

1 L ( yi , f ( xi ) ) l i =1

µ ⎛ ⎞ ∑+ ⎜ g j − µ+ −+µ− ⎟ ⎠ j∈S ⎝

the degree of scatter of negative sample class s−2 :

N F ( w ) said that the probability of the samples were misclassified when the direction of decision hyperplane is w. The definition of the risk of experience [1][2] Re mp ( w ) =

(9)

J F ( w) =

sb2 s+2

+ s−2

=

1 s+2

+ s−2

(13)

Involved in calculating the variables of J F ( w ) are directly associated with the w, while w depends on the selection of kernel function. In other words, reflects the kernel function for image set has the influence on the degree of scatter within class and the degree of scatter between class in the direction of

The Degree of Scatter of Sample Class

LF ( w ) only characterizes the degree of linear separation of f , but cannot reflects the distance between

and within the classes of f . Therefore introduce some of ideas of the degree of scatter of sample class in Fisher Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


991


w straight line projection. SVM training aims to maximize the interval | w | , the greater | w | the greater the confidence interval. And the smaller the degree of intensity between sample class，the greater the training received | w | . Call the reciprocal of J F ( w ) as:

V.

The experiment selected the Adult data set a1a and a5a in University of California, Irvine (UCI) to test. Adult data set has been widely used to test the classification algorithm, after discretized the data, the data set for each sample has 0-1 characteristics as number as 123. CPU hardware platform for the experimental use of AMD2600, memory 512M, the operating system using Windows 2000, develop experiment is MS Visual C++6.0. The simulation results shown in Tables I, II. The experimental results confirmed the validity of the method described in this paper, the more a5a samples, the more its distribution similar to the normal distribution.. The result shows, the decision function depends on the balance between N F ( w ) and S F ( w ) , which also illustrate the method line with the principle of structural risk minimization method.

Definition 2 The degree of intensity between sample class S F ( w ) : SF ( w) =

1 = s+2 + s−2 J F ( w)

(14)

The larger S F ( w ) ,the greater the interval | w | .Similar to the definition of expected risk,the performance of kernel function is the sum of non-linearity of sample and the degree of scatter between sample class.

TABLE I EXPERIMENTAL RESULTS(1) ROE model NAPS model -85.876** -22.056** (-4.376) (-3.456) -0.09** (-4.969)

Variable Constant DAR DAR2 SKRCG G5 GYFR YH LNA Ind F-statistics Adjusted R2 Sample number

Simulation

ROA model -43.726** (-4.282) -0.164**(-4.928)

0.256**(4.255)

0.085** (3.082)

0.125**(4.533)

3.934**(4.356)

1.297**(4.109)

2.199**(4.47)

20.406** 0.249 118

10.527** 0.246 118

19.435** 0.321 118

TABLE II EXPERIMENTAL RESULTS(2) Data set

Training samples / Test samples

Kernel

Polynomial kernel a1a

1324/40783 RBF kernel

Polynomial kernel a5a

4420/34870 RBF kernel

VI.

Parameters

N F ( w)

SF ( w)

d=0.6 d =1.2 d =1.8 g =0.020 g =0.080 g =0.04 d=0.6 d =1.2 d =1.8 g =0.020 g =0.080 g =0.04

0.14 0.02 0.01 0.25 0.17 0.16 0.15 0.08 0.02 0.13 0.16 0.16

15.625 7.1418 1.1815 14.4927 13.6986 12.6582 32.2581 25.6410 10.5263 26.3157 27.7778 23.8095

Training samples Test samples Correct classification Correct classification rate rate 70% 80% 85% 81% 89% 81% 80% 79% 90% 80% 95% 79% 92% 88% 91% 90% 97% 78% 86% 82% 82% 85% 82% 84%

interval, future work will start from these two angles, to research quantitative analysis of penalty factor C in the standard SVM.

Conclusion

This paper from the perspective of probability and statistics，quantitative analyzed the influence of kernel function to decision-making function by nonlinearity and dispersion of samples. Experimental results show that the method is effective, leading a way for the selection of kernel function. As the nonlinearity and dispersion of samples reflect the risk of experience and confidence

Acknowledgements The authors are grateful to the anonymous reviewers and to the Natural Science Foundation and Scientific



992


Research Fund of Hunan Provincial Education Department (11B031). Authors gratefully acknowledge the Projects Supported by Scientific Research Fund of Hunan Provincial Education Department (09C271 &11C0327&10C0569) for supporting this research. Project supported by Provincial Natural Science Foundation of Hunan（10JJ6099&11JJ9019）supports the research. Project supported by Provincial Science & Technology plan project of Hunan (2010GK3048) supports the research. This research is supported by the construct program of the key discipline in Hunan province.


Post-Doctoral Research Center, Guanghua School of Management, Peking University Beijing, P.R.China, 100872. 2 Baoshang Bank Post-Doctoral Research Center, Beijing, P.R.China,100101. 3

School of Information Engineering, Henan Institute of Science and Technology. Xinxiang, Henan, China, 411104. Wenfang Sun, graduated from Southwest Jiaotong University, and obtained Ph.D. in Major of Logistics Engineering in 2010. In the international journals,she has published 6 papers indexed by EI, and published 2 papers in Chinese core journal.She is a postdoctor in Baoshang Bank and Peking University now, and her research interesting is supply chain finance.

Reference [1] [2] [3] [4] [5]

[6]

[7]

[8]

[9]

[10] [11]

[12]

[13] [14] [15] [16]

[17]

[18]

http://www.demica.com/solutions/supply-chain-finance Vladimir N.Vapnik. The nature of statistical learning theory. (Tsinghua University Press),2000. Vladimir N.Vapnik. Statistical Learning Theory. (Publishing House of Electronics Industry),2004. Dengnai Yang, Yingjie Tian. New Method for Data Mining-SVM[M] . (Science Press),2004. Mika S, Ratsch G, Weston J, Scholkopf B, Müller K R. Fish discriminateanalysis with kernels. In.Hu Y H, Larsen J, Wilson E, Douglas S, eds. Neural Networks for Signal Processing IX, IEEE. 1999. pp:41-48 Rosipal R, Trejo L J, Cichocki A. Kernel Principal Component Regression with EM Approach to Nonlinear Principal Components Extraction. Technical Report No.12, Department of Computing and Information Systems, University of Paisley, 2000.11 Francis R Bach, Michael I Jordan. Kernel Independent Component Analysis. Journal of Machine Learning Research (S1533-7928). 2002,3. pp:1-48. Ling Zhang. SVM based on kernel function and forward neural network with three relations . Chinese Journal of Computers, 2002, 25(7), pp:696-700 Jingtao Huang, Longhua Ma, Jixin Qian. A Classification Problems For Multi Improved SVM . Journal of Zhejiang University, 2004.12 Zhiquan Qi, Yingjie Tian, Zhijie Xu. Kernel Parameter Selection In SVM . Control Engineering, 2005.12(4), pp:379-381 Chunxi Dong, Xian Rao, Shaoquan Yang, Songtao Xu. SVM Parameters Selection Method. Systems Engineering and Electronics, 2004.26(8) pp:1117-1120 Drucker H, Wu H D. Support vector machines for spam categorization. IEEE Transaction on Neural Network, 1999.10(5), pp:1048-1054. Thorsten Joachims. Learning to classify text using support vector machines. Dissertation, Universitaet Dortmund, February 2001. Ju Sheng etc. Probability and Mathematical Statistics . Higher Education Press, 1999 Fisher R A. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 1936, 7(2), pp:179-188. Xilong Qu, Linfeng Bai, Xiaobo Ming, Defa Hu. Anti-Collusion Digital Fingerprinting Based on IPP Code, International Review on Computers and Software (IRECOS),Vol.6,No.5,pp:534-541,2011. Xilong Qu, Zhongxiao Hao,Linfeng Bai, Xizheng Zhang, Design of T-S Fuzzy Control for Active Queue Management in TCP Networks. International Review on Computers and Software (IRECOS),Vol.6, No.5, pp:534-541, 2011. Xilong Qu, Zhongxiao Hao, Xizheng Zhang. A Novel Adaptive Sliding Mode Congestion Control Algorithm for TCP Networks via T-S Model Approach. International Review on Computers and Software (IRECOS), Vol.6, No.5, pp:534-541,2011.

Jindong Zhang, graduated from Nankai University, in economics. and obtained Ph.D. in Economics. Now, he is mainly engaged in banking operations management research.

Xilong Qu, PH.D., associate professor of Hunan Institute of Engineering, master supervisor of Xiangtan University, the key young teacher of Hunan province, academic leader of computer application technology in Hunan Institute of Engineering. His research interesting is web service technology, information safety and networked manufacturing. He has presided the Projects Supported by Scientific Research Fund of Hunan Provincial Education Department(08A009) , the projects supported by Provincial Natural Science Foundation of Hunan（10JJ6099 & 11JJ9019）and the project supported by Provincial Science & Technology plan project of Hunan(2010GK3048). He has published more than 30 papers in some important magazines.



993


Leveraging Historical Assessment Records for Constructing Concept Maps M. Al-Sarem1, M. Bellafkih2, M. Ramdani1 Abstract – In recent years, researchers have proposed different approaches for developing adaptive learning systems based on learning behaviors of learners during their interaction with elearning systems. For achieving the adaptive learning, a predefined concept map of a course is often used to provide adaptive learning guidance for learners. However, it is difficult and time consuming to create the concept map of a course. In this paper, we apply Data mining techniques to constructing concept maps based on the historical testing records. The paper, first, provides a common basis for the analysis of automatic concept maps building approaches meeting in literatures. Then, the general process of automatically constructing concept maps using historical assessment records is proposed. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Association Rules, Concept Map, Domain Model Representation, Educational Data Mining, Formal Student Model

Nowadays educational data mining is one of rapidly growing research community, especially with innovation of new technologies as web-based education, where the data collected during the learning process can be contain a huge mass of hidden data which can be useful for mining the hidden relationships [1]. The web-based education systems normally record the student’s accesses in web logs that can provide a raw trace of the learners’ navigation on the site [2]. According to Srivastava et al. [3], there are three types of logs: 1. Server log file which constitutes the most widely used data source such as learning path, time and inputresponse. 2. Client log file, which consists of a set of log files, one per student, and contains information about the interaction of the user with the system. 3. Proxy log file, which consists of a set of log files of caching between client browsers and web servers. This information complements server log file information. However, there are many restrictions related to log files (log files have several limitations such as tracking for files not users, simple click and not learning activities and having incomplete and incorrect information problem [2]). To cope with this problem, researchers have proposed several solutions: combing usage data, content data and structure data in a web site to generate user navigational models[4],expand automatically generated log files by introducing contextual information as additional events and by associating comments and static files[5], combing data with other inquiry methods,

such as informal chatting with students[6] or data on the activity with content and user profiles in a composite information model[7]. Collected data in web-based education system can be distinguished by their types: 1. Data that describe the pattern of usage of web pages, such as information about the student’s actions and communications, or information about the student’s activities in the course. 2. Data that provide personal information about users of the web site, such as registration data and user profile information. We also distinguish between three different types of web-based education systems [2]: 1. Particular web-based courses which use standard HTML to navigate their contents. 2. Well-known learning content management systems that offer a great variety of channels and workspaces to facilitate information sharing and communication between participations in a course. 3. Adaptive and intelligent web-based educational systems that attempt to be more adaptive by building student model and using this model throughout the interaction with the student in order to adapt to the needs of that student. From this end, the data from Adaptive and intelligent web-based educational are semantically richer than and can lead to more diagnostic analysis than data from traditional web-based education system [8].Moreover, available data can be collected from the domain model, pedagogical dataset (set of problems and their answer) and student model[2]. The aim of this paper is to construct domain model as concept maps based on the historical assessment records stored in student model.



I.

Introduction

994

M. Al-Sarem, M. Bellafkih, M. Ramdani

A view of perspectives and respective classes are depicted by Fig. 1. Along this section, we define each one of them.

As known, historical testing records can be useful for constructing concept maps [9] of course, thus, applying data mining techniques are a powerful tool in discovering hidden knowledge from the testing results [10]. In recent years, researchers have begun to investigate various data mining methods in order to help learners improve their learning process. Some of the most useful data mining tasks and methods are clustering, classification and association rule mining for example, in the National University of Singapore the data mining application have been used for classifying and selecting those students who need extra classes in a given subject. With the help of data mining they are able to select the targeted students much more precisely than by traditional methods [11]. Myller et al., in [12] applied EM (Expectation-Maximization) – algorithm for clustering the students to construct homogeneous groups in terms of programming skills according to the students’ skills, to predate exam results according to the skills shown in exercises. Behrooz & Barnes [13] suggested to use educational data mining methods as a way for creation a system that can judge a student’s performance by the way he/she responds to questions (they gathered data from an introductory programming course teaching C++) for determining where the student needs to help. Dominguez et al. [14] integrated data mining into an e-learning system to generate dynamically tailored hints for students who are completing programming exercises during a national programming online tutorial and competition. Return to construct concept maps of domain model, Bai and Chen [15] applied fuzzy rules and fuzzy reasoning techniques to automatically concept maps based using students’ historical testing records. Tsai et al. [16] presented a method for a two-phase fuzzy mining and learning algorithm for an adaptive learning environment. Sue et al. [17] presented a two-phase concept map construction algorithm where the fuzzy association rules have found by mining the students’ answers in adaptive testing systems. A different approach has proposed by Lee et al. [18], where process of constructing concepts maps for conceptual diagnosis based on Apriori algorithm [19]. In next section, we provide a review of concept maps building approaches.

II.

II.1.

Nature of Data Sources

The nature of data source used by an approach impacts the whole concept map building process [20]. In educational learning environments, specifically in assessment system, learner’s grades can be sorted in binary or numerical formats. From this standpoint, data sources used by the approaches could be classified as binary and numerical data sources. Among the approaches that use binary data sources, we could mention [21]-[24] and those that use numeric data sources [16], [25], [26].

Fig. 1. Classes for Analyzing the Approaches to Build Concept Maps

II.2.

Methods and Outcomes Perspective

Based on the nature of the collected data, as we saw in the previous section, several approaches have been proposed: For binary data, a classical data mining methods are used [23]. However, for numerical data sources, a combination of fuzzy set theory and data mining techniques was proposed a clear example found in [16] where an association rule mining algorithm, called Apriori, has integrated with fuzzy set theory to find embedded information that could be fed back to teachers for refining or reorganizing the teaching materials and tests. In a second phase, an inductive learning algorithm of the AQ family has been used: AQR, to find the concept descriptions indicating the missing concepts during the students’ learning. The results of this phase could also be fed back to teachers for refining or reorganizing the learning path. The same approach, we can find in [27], where Huang proposed an adaptive tutoring learning model to find out all the prerequisite relationship between concepts, and to construct a concept map. According to graphical layout of a concept map, we may classify these approaches in tree or graphic classes. If there is an explicit topology, the concept map belongs to a tree class, otherwise, a graph class.

A Review of Concept Maps Building Approaches

In this section, we propose as a common basis for the analysis of automatic concept maps building approaches meeting in literatures. We organized these classes according to different perspectives: 1. Nature of data sources: What kind of data source was used? Where we can collect it and techniques were used to handle it? 2. Using methods: What methods used to construct concept maps? 3. Outputs: How the resultant output looks like? Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved


995


={ , ,

III. General Process of Automatically Constructing Concept Maps Using Historical Assessment Records

where: - Testing question; R- Student’s answer; M- Student’s mark. , where, ,

Analysis shows that all the mentioned above works have four phases, including the Data preprocessing, the Mining prerequisite relationships Concept Map constructing process and Adjustment final map, as shown in Fig. 2.

}

;

;

Maximum marks can be obtained by student , where, , Student’s getting mark and 0 . Having that, the grade matrix based on the proposed student model will be shown as follows: … … …

Fig. 2. General concept maps constructing process based on historical assessment records

III.1. Data Preprocessing

…

Data preprocessing allows transforming the original data into a suitable shape to be used by a particular mining algorithm. So, before applying the data mining algorithm, a number of sub-phases data preprocessing tasks have to be addressed: • Data preparation process. As most of web-based educational systems record students’ grades directly in student model, which is stored in databases. Thus, data must prepare to be adopting for the mining process. • Data transformation. The aforementioned algorithms such as in [1], [22] for mining constructing concept map deal with binary/ numeric data that do not ready to use. Thus, dependant on nature of data, either using students’ grades directly or need to fuzzificate it. • Data reduction: It is for reducing data dimensionality. We propose three ways to reducing the input data before the mining process: Clustering questions by their difficulties, using an anomaly diagnosis process [25] and Item Response Theory Data Preprocessing Approach [28]. According to Bule [29], the formal student model for adaptive learning system consists of the following components: ,

,

,

where “ , ” denotes the score of question “ ” of the learner “ ”, , = 1 denotes the student “S ” gets the right answer in question “ ”,and = 0 denotes the , student “ ” has a wrong answer in question “ ” if collected data have been stored in binary data or , [0, ], 1 and 1 , “n” is number of learners and “m” is number of questions. III.2. Mining Relationships Process In the adaptive learning environment, the learning status of a concept can possibly be influenced by learning status of other concepts. Appleby et al. [30] proposed an approach to create the potential links among skills in math domain, where the direction of the link is determined by a combination of educational judgment, the relative difficulty of the skills, and the relative values of cross-frequencies. From pedagogical view of point, a harder skill should not be linked forwards to an easier skill. Let represents the amount of learners with wrong answers of skill and right answers of skill . , a skill A could be linked to a harder If skill B, but backward link is not permitted. The same approach is correct for any two concepts found in prerequisite relationship, if is the and prerequisite for efficiently learning the more complex and higher level concept , then a relationship exists. Using grades matrix , the mining relationship process formulate as following: Let the test portfolio ,…, be a set of questions tested by students , … , . Find all , where: association rules 1. For Binary data stored in student model, find all is correctly association rules, where the question

,

where: – General information about student; – Background knowledge; – Work with a course (a system); – learning method or/and strategy; – Psychological characteristics. For effectively leverage data stored in aforementioned has the following formulation: student model,



996


relationship map) based on the heuristics shown in Table I: • For the association rules, where question to influences question , we add an edge from into the question-question relationship map. • For the association rules, where question influences question , we add an edge from to into the question-question relationship map. Let the confidence of an association rule be the confidence of the relationship between questions builds from it and for any two questions, if the confidence of the questions-relationship is smaller than the minimum confidence “θ” given by teacher/ expert, then we delete the relationship between them to get the completed question-relationship map. If there is more than one relationship between any two questions, then we only keep the relationship with the maximum degree and delete the others. For all association rules “ ” obtained in the previous Step, we calculate the relevant degree “ ” between concepts “ ” and “ ” from the relationship “ ”, as follows [21]:

learned by the learner, then the question is also correctly learned . Also for case, where question is incorrectly learned by the same is also incorrectly learner, then the question . learned 2. For numeric data, find all association rules, where “ , , ”, “ , , ” and “ , , ” (the related explanations of the analysis are shown in Table 1) As known, the numeric testing data are hard to analyze by association rule mining approach. Therefore, for mining relationships types “ , , ”, “ , , ” and “ , , ”), fuzzy set theory can be useful to transform these into symbolic. Thus, if the membership functions of each quiz’s grade are known (see Fig. 3) and we fuzzificated the grades matrix to symbolic data, then the Look Ahead Fuzzy Association Rule Mining Algorithm (LFMAlg) [16] can be applied to mine some rules of test items, which are used to construct the concept maps and fed back to teachers for further analyzing.

rev ( Ci → C j )

Qx Qy

where “ of the relationship “

In the fuzzification result, “LOW”, “MIDDLE” and “HIGH” denote “a learner " " has a low grade in question “ ”, a learner " " has a middle grade in question and a learner " " has a high grade in question respectively. TABLE I THE EXPLANATIONS OF RULE TYPES [25]

Q ,L

Q ,L

Q ,H

Q ,H

Q ,L

Q ,H

Q ,H

” denotes the relevance degree " converted from the

… … …

Description of relationships

Q ,L

(1)

”, [0,1], “C ” relationship “ denotes a concept appearing in the question “ ”, “ ” denotes a concept appearing in the question ” denotes the weight of the concept “ ” in “ ”,“ ” denotes the weight of the the question “ ” ,“ ” concept “ ” in the question “ ”, “ denotes the confidence of the relationship “ ”, “ ” and ,“ ” can be obtain from questionconcept matrix, x ≠ y, 1≤ x ≤ m, 1≤ y ≤m and 1 . The questions-concepts matrix QC shown as follows:

Fig. 3. The given membership functions of each quiz’s grade

Rule

= WQx Ci × conf ( Qx → Q y )

It is means that the related concepts in question “ Q ” are the prerequisite of those in “Q ” and explain why getting low grade in question “Q ” might imply getting low grade on “ Q ” It is means that the related concepts in question “Qj” are the prerequisite of those in“Qi” because “Qi" may be not learned well resulting from “Qj” It is means that the concepts in question “Qi" are the prerequisite of concepts in “Qj” It is means that the concepts in question “Qi" are the prerequisite of concepts in “Qj”

… where: " " - denotes the degree of relevance of question with respect to concept (" 1" if a test question contains a single concept, [0, 1] if a test question contains more than one concept and " 0" if does not contain any concept, moreover, the total weight of the test question is 1). ” be the confidence Furthermore, let “ of the relationship “ ”. If there is more than one relationship between any two constructed concepts, then

III.3. Concept map Constructing Process Before constructing the concept map, we propose to construct the casual effect map (Question-Question



997


0,

the relationship between the two concepts chosen as follows:

,

where " " denotes the minimum nonzero values in a concept ’s column and ’s column for all questions from matrix QC, [0, 1] and 1 ≤ i, j ≤ n. The aim of using concept-concept matrix is to find out the minimum nonzero value of the relevance degree in C' . matrix For each relationship " ", calculate the relative questions-concepts value between concepts shown as follows:

(2) III.4. Concept Map Adjustment Process The proposed algorithm above causes sometimes unreasonable prerequisite relationships of concept test, where two concepts can be prerequisite each other. Thus, we proposed three criteria to cope with this problem: 1. Using confidence and the Interest factor measure (IF measure) to detect circularity in question- relationship maps; 2. Using the relevance degree between any two concepts to de detect whether a cycle exists or not; 3. Calculating the relative values of each two concepts to keep relationship between it.

(3) where: - The relative questions-concepts value between concepts ; - The number of the questions which have concept ; - The number of the questions which have concept . The relative questions-concepts value measures the weight of concepts in each question that support minimum of relevance degree of question with respect to concept :

III.4.1. Cycle Detection Process We detect the circularity firstly in the constructing question-question relationship map. Using confidence and IF measure are useful to detect the redundancy and circularity as it can filter many useless rules. For the relationships, where the concept and concept influence each other, e.g., the relevance degree can be also criteria to choose the higher effect relationship. Rule: keep the relationship and If delete the other.

μ

Conclusion

Historical testing records can be contain a huge mass of hidden data which can be useful for constructing concept maps of course, thus, applying the fuzzy set theory and data mining techniques are a powerful tool in discovering hidden knowledge from the testing results, therefore, many researchers have proposed various approaches and methods for developing adaptive learning systems based on concept maps. In this paper, we have presented an innovation method to automatically construct concept maps based on historical students’ grades. The general process for automatically constructing concepts maps based on students’ grades have four phases, including the Data preprocessing, the Mining prerequisite relationships Concept Map constructing process and Adjustment final map. As, most of web-based educational systems record students’ grades directly in student model, which is stored in databases, we need to adopt stored data in

Fig. 4. Detection of Circularity

III.4.2. Adjustment Process To use a relative value of each two concepts as criteria for keeping the relationship, we have to create a new concept-concept matrix C' based on the conceptual weights found in matrix QC, shown as follows:

′

(4)

If Eq. (4) is true and the relevance degree between concepts is greater than " ", we consider the relationship ” is important and add an edge from to “" into the concept map with the relevance degree of relationship " " to construct a concept map. Otherwise, delete it.

IV.

0

,

… … 0 … 0



998


[11] Ma, Y., Lee, S.M., Liu, B., Yu, P.S., Wong, C.K., Targeting the Right Students Using Data Mining, Proceedings of the Sixth ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (p. 457-464 Year of Publication: 2000). [12] Myller, N., Suhonen, J.,Sutinen, E., Using Data Mining for Improving Web-Based Course Design, Proceeding of the International Conference on Computers in Education (ICCE'2) (2002). [13] B. Mostafavi, T. Barnes, Towards the Creation of a Data-Driven Programming Tutor. Intelligent Tutoring Systems, 2, 2010, 239241. [14] A. K. Dominguez, K. Yacef, J. R. Curran, Data Mining to Generate Individualised Feedback. Iintelligent Tutoring Systems, 2, 2010, 303-305. [15] S.M. Bai, S.M. Chen, Automatically constructing concept maps based on fuzzy rules for adapting learning systems, Expert systems with applications, Vol. 35, pp. 41-49, 2008. [16] Tsai, C. J., Tseng, S. S., & Lin, C. Y., A two-phase fuzzy mining and learning algorithm for adaptive learning environment, In Proceedings of the international conference on computational science , Lecture notes in computer science (LNCS 2074), Vol. 2, Page: 429–438 Year of Publication: 2001). [17] Sue, P. C., Weng, J. F., Su, J. M.,Tseng, S. S., A New Approach For Constructing the Concept Map, In Proceedings of the 2004 IEEE international conference on advanced learning technologies (Page: 76–80 Year of Publication: 2004). [18] C. H. Lee, G. G. Lee, and Y. H. Leu, “Application of automatically constructed concept map of learning to conceptual diagnosis of e-learning”, Accepted and to appear in Expert Systems with Applications, 2008. [19] Agrawal, R., Srikant,R., Fast algorithms for mining association rules, Proceedings of the 20th International Conference on Very Large Database (Page: 487—499 Year of Publication: 1994) [20] Kowata, J. H., Davidson, C. M., Claudia, S. B., A Review Of Semi-Automatic Approaches To Build Concept Maps. Concept Maps: Making Learning Meaningful, Proceedings of Fourth International Conference on Concept Mapping, 2010. [21] C. H. Lee, G.G. Lee, Y. Leu, Application of automatically constructed concept map of learning to conceptual diagnosis of elearning. Expert Systems with Applications, Vol. 36, n.2, pp.16751684, 2009. [22] M. AL-Sarem, M. Bellafkih, and M. Ramdani (2011). A New Method for Constructing Concept Maps in Adaptive E-Learning Systems, Springer-Verlag Berlin Heidelberg. CESM 2011, Part II, CCIS 176, pp. 180–185. [23] Bai, S.M., Chen,S. M., A New Method for Automatically Constructing Concept Maps Based on Data Mining Techniques, Proceedings of the Seventh International Conference on Machine Learning and Cybernetics,2008. [24] G.J. Hwang, Judy C. R. Tseng, and G.H. Hwang, Diagnosing Student Learning Problems Based on Historical Assessment Records, Innovations in Education and Teaching, Vol. 45, n.1, pp.77 – 89, 2008. [25] S.S. Tseng, .S. P.C. Sue, J.M. Su, J.F. Weng, W.N. Tsai, A new approach for constructing the concept map, Computers & Education, Vol. 49, pp. 691–707, 2007. [26] S.Y. Cheng, C.S. Lin, H. H. Chen, J. S. Heh, Learning and Diagnosis of Individual and Class Conceptual Perspectives: an Intelligent Systems Approach Using Clustering Techniques, Computers & Education, Vol. 44, n. 3, pp.257–283, 2005. [27] Hung, C. L., Hung, Y. W., A Practical Approach for Constructing an Adaptive Tutoring Model Based on Concept Map, VECIMS 2009 - International Conference on Virtual Environments, Human-Computer Interfaces and Measurements Systems (Page 298 – 303 Year of Publication: 2009). [28] C. Y. Liao, Approach for Constructing the Concept Effect Relation Map of Mathematics, MS. Thesis, National Chiao Tung University, Hsinchu, Taiwan, 2005. [29] Bule, J., Adaptive Computer-aided Teaching Methods based on Student Mode, Proceeding of First International Conference “Information Technologies in Education for All” (Page: 221– 230 Year of Publication: 2006).

student model. Thus, data must prepare to be adopting for the mining process, we presented a formal student model which distinguishes data by their types. For effectively mining relationships, we applied the association rules techniques to mine the prerequisite relationships, where Apriori algorithm used to mine rules from data stored in binary format and Look Ahead Fuzzy Association Rule Algorithm for numeric data. In constructing phase, two kinds of questionsrelationship maps have been constructed. After that, both maps combined into the one combined questionsrelationship map. This step is necessary to minimize the count of operations and to economize the space need to store the result of calculations. Also, we have calculated the relevance degree between any two concepts to construct maps. As the proposed algorithm above causes sometimes unreasonable prerequisite relationships of concept test, where two concepts can be prerequisite each other. Thus, we proposed three criteria to cope with this problem: using confidence and the Interest factor measure (IF measure) to detect circularity in question- relationship maps, using the relevance degree between any two concepts to de detect whether a cycle exists or not and calculating the relative values of each two concepts to keep relationship between it. The proposed method provides us with a useful way to construct concept maps in adaptive learning systems, moreover, can be useful tool to represent domain model as semantic network.

References [1]

M. Al-Sarem, M.Bellafkih, M. Ramdani, Mining Concepts’ Relationship Based on Numeric Grades. IJCSI. Vol. 8(4), n. 2, pp.136-142. 2011. [2] C.Romero, S. Ventura, Educational Data Mining: A Survey from 1995 to 2005, Expert systems with Applications, Vol. 33, pp. 135146, 2007. [3] J., Srivastava, R., Cooley, M., Deshpande, P., Tan, Web Usage Mining: Discovery and Applications of Usage Patterns from Web Data. SIGKDD explorations, Vol. 1, n. 2, pp.12-23, 2000. [4] Li, J., Zaïane, O., Combining usage, content, and structure data to improve web site recommendation, In International Conference On E-Commerce And Web Technologies. pp. 305-315, 2004. [5] Avouris,N., Komis, V., Firotakis, G., Margaritis, M., Voyiatzaki, E., Why logging of fingertip actions is not enough for analysis of learning activities, In Workshop on usage analysis in learning systems at the 12th international conference on artificial intelligence in education, 2005. [6] A. Ingram, Using Web Server Logs in Evaluating Instructional Web Sites. Journal of Educational Technology System. Vol. 28, n. 2, pp. 137-157, 1999. [7] D. Monk, Using Data Mining for e-Learning Decision Making. Electronic Journal of e-learning, Vol. 3, n. 1, pp. 41-54, 2005. [8] A. Merceron, K. Yacef, Mining Student Data Captured from a Web-Based Tutoring Tool: Initial Exploration and Results. Journal of Interactive Learning Research, Vol. 15, n. 4, pp. 319346, 2004. [9] J.D. Novak, Learning, Creating, and Using Knowledge: Concept Maps as Facilitative Tools in Schools and Corporations. NJ: Lawrence Erlbaum Associates, 1998. [10] Liao, C. Y., Tseng, S. S., Weng, J. F., An IRT-Based Approach to Obtaining Item-Aware Learning Achievement, Proceedings of the 23th workshop on combinatorial mathematics and Computation Theory,( p.362-368 Year of Publication: 2006).



999


[30] J. Appleby, P., Samuels, T. T. Jones, Diagnosis – a KnowledgeBased Diagnostic Test of Basic Mathematical Skills, Computers and Education, Vol. 28, n. 2, pp. 113-131, 1997. [31] M. H. Marghny, A. A. Shakour, Fast, Simple and Memory Efficient Algorithm for Mining Association Rules, International Review on Computers and Software (IRECOS), Vol. 2. n. 1, pp. 55 – 63, 2007.


Département d’informatique, FSTM, Mohammedia, Morocco. 2 Institut National des Postes et Télécommunications (INPT), Rabat, Morocco. Mohammed Al-Sarem received: The Research Master's in Computer Engineering and Information Technology from Volgograd State Technical University, Russia in 2007, and a BSc. in Computer Science in 2005 from the same university. He is conducting, since 2010, doctoral research in Computer Science in University of Hassan II, Mohammadia, Morocco, at e-learning systems: Models and methods of providing intellectual adaptive learning guidance in elearning environment. Mostafa Bellafkih received: PhD thesis in Computer Science from the University of Paris 6, France, in June 1994 and Doctorate Es Science in Computer Science (option networks) from the University of Mohammed V in Rabat, Morocco, in May 2001. His research interests include the network management, knowledge management, A.I., Data mining and Database. He is Professor in The National Institute of Posts and Telecommunications (INPT) in Rabat, Morocco since 1995. Mohammed Ramdani received: The PhD thesis in Computer Science from the University of Paris 6, France, in February 1994 and Habilitation in Computer Science from the University of Paris 6, France, in June 2001. His research interests include the, knowledge management, A.I., Data mining and Database. He is Professor in Mohammedia Faculty of Sciences and Technologies (FSTM), Morocco since 1995.



1000


Object Oriented Database Applying Study on ISO 9001:2000 System K. Khoualdi1, T. Alghamdi2

Abstract – Database model is an essential topic in software engineering. This paper highlights the technical and commercial differences between object oriented database management system OODBMS and relational database management system RDBMS and follows up some of these differences on a part of ISO 9001:2000 system's database. We apply both models, object oriented database model OODBM and relational database model RDBM, on the same part of ISO 9001:2000 system's database to study how each model represents and accesses the data. This paper provides comprehensive information about OODB and specially guides the students and worker in software engineering to select database model which is matching with their application's requirements. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: ISO 9001:2000, Object Oriented Database Model, Relational Database Model

I.

Introduction

"The choice of an appropriate representation for the structure of a problem is perhaps the most important component of its solution. For database design, the means of representation is provided by the data model" [1]. The common database models are relational model and object oriented model. Relational model is the second generation of Database Management System DBMS. It was proposed in 1970 by Edgar Codd to represent the natural structure of data by using a table, where each row in the table refers to an entity and each column refers to an attribute of the entity, With this model the user does not need to know about the machine representation. The user can retrieve the information from one table or more by using a high-level nonprocedural language SQL. Relational Database Model RDBM made a huge leap in the field of DBMS. It used the idea of end-user programming and interactive querying of a database [2] – [3]. Object model is the third generation of DBMS. It started at the beginning of 90, but it initially came out from academic research in the middle of 90. It started in the engineering and design field, and it became the favored system for financial and telecommunications applications. There were high expectations for Object Oriented Database Management System OODBMS in that would become the primary database technology, instead of Relational Database Management System RDBMS, but none of these predictions were achieved. RDBMS are still by far the most widely used databases. In fact this failure is referring to technical and commercial reasons.


1001

The technical reasons can be summarized by the developments which RDBMS made, like added object capabilities Object Relational Database Management System ORDBMS, improved optimizers, and implemented Standardization. The commercial reasons can be summarized by the difference between OODBMS and RDBMS in their goals and by their ages. OODBMS was created to find a solution for complex software, where RDBMS focused on business applications which represent a large market. RDBMS has a bigger installed base, and RDBMS vendors have more money and marketing share than OODBMS vendors [4].

II.

Object Oriented Concept II.1.

Inheritance

Classes can inherit the attributes and behaviors of other classes. They represent a hierarchy structure where the higher class which generates other classes called (super classes or parent), and the lower class called (subclasses or child). Subclass can inherit from one class (Single inheritance) or more classes (Multiple inheritance). New subclasses can override attributes and methods inherited from super classes. Inheritance develops complex software by creating new objects and allowing it to hold new features in addition to all old features. Benefits of inheritance can be summarized on malleability and reusability, malleability means a set of objects sharing common behavior, this facilitates program construction, maintenance, and extension; reusability means reuse the code and data just by writing new abstractions [5].


K. Khoualdi, T. Alghamdi

II.2.

Polymorphism

III.2. Data Model

Much power of object-oriented approach derives from the mechanism of polymorphism, it allows sharing specification of the operation with other objects, these objects can further extend this operation to provide behaviors that are unique to those objects, but the problem is that both of the operations have the same name. The Interface can be separated from the implementation, so to select a particular implementation of a given operation one must be familiar with at least one of these implementations [6]. II.3.

Encapsulation

The manipulation of an object is possible only through its defined external interface, the instance variables and methods are hidden, and as a result, the implementation can be changed without affecting in the code that uses the object’s interface. Encapsulation is proper for many software applications, it plays a unique role, because it is able to directly reflect the barriers in the organization domain, but in some situations encapsulation is not desirable. In fact, programming language designed to completely support Encapsulation or not, Stephan Herrmann defines gradual encapsulation as an approach where technology supports both encapsulation and its absence [7].

III. Comparison of OODBMS to RDBMS The different between OODBMS and RDBMS goes back to the different foundations. The relational model is based on the mathematical concept of sets, where the object oriented model is based on the object-orientation mechanism. III.1. Data RDBMS uses tables to represent data. This structure is inefficient, leading to many joints during query processing and poor representation of real world entities [8]. OODBMS uses Object to represent data. It can represent real world and complex relationships, and it can represent a hierarchical structure. It is also able to develop systems faster than RDBMSs by using inheritance. Object structure supports Encapsulation, concurrency, and ad hoc query [9]. RDBMS just store native data type like integers, floating point, character, strings, date-time, and currency, where OODBMS can store more data like images, video, audio, animations and mixed media. Accessing data in RDBMS depends on joints between tables through the query process, where accessing data in OODBMS can be faster because objects can be retrieved directly by following pointers.


Data models represent entities and their relationships, constraints and operations that change the states of the data in the system. Application with RDB usually needs more than one model, an Entity Relationship diagram to model the static parts and a separate model for the operations and behaviors of entities, where application with OOD can be modeled in one UML diagram [10]. III.3. Relationships RDBMS use tables to represent relationships; it does not distinguish between entities and relationships, or between different types of relationships [8]. OODBMS offer the feature of inverse relationships, and it does not need to define any key. III.4. Unique Identifiers UIDs RDBMS supports only unique identifiers by userdefined, this mechanism has a natural, human readable meaning, it is good in the situation where primary keys exists for a collection that is known never to change, Consequently, in data interchange or debugging this may be an advantage. OODBMS supports an OID that is automatically generated by the system. This mechanism guarantees uniqueness to each object, an OID cannot be modified by the application, and this way eliminates the need for user defined keys in the OODB model [11]. Also, equality in OODBMS between two objects depends on their values or OIDs, where in RDBMS equality is always based only on values [12]. III.5. Business rules RDBMS supports referential integrity, domains, and no more business rules. In OODBMS programmers can add any rules by writing codes in one or more functions, but with associating rules with functions it becomes very hard to create new rules, update or delete old ones. Moreover, there is no easy way to query the rules [11]. III.6. Platform RDBMS already supports codes in an object oriented programming language, but using OOPL with RDBMSs leads to impedance mismatch, although there are some proposals known as native queries that support some kinds of queries within typical programming languages, these approaches still face limitations [13]. Some object-oriented databases are designed to work well with object-oriented programming languages such as Java, C#, Visual Basic .NET, C++ and Smalltalk, while others have their own programming languages which remove the impedance mismatch, but it has restrictive choices of programming languages. International Review on Computers and Software, Vol. 6, N. 6

1002


III.7. Standardization RDBMS uses SQL standards, which works as interface between the user and the database. It was adopted by the International Organization for Standardization ISO and the American National Standards Institute ANSI. In 1993 The Object Database Management Group ODMG developed standards for object-database and object-relational-mapping products. The group created an Object Query Language standard OQL, but very few database vendors implemented it [4]. III.8. Encapsulation and Inheritance There are two benefits of encapsulation: less programming complexity and simpler modifications, but at the same time encapsulation effects on query, the optimizer needs greater freedom for devising an efficient plan. Thus encapsulation boosts resilience but limits optimization potential, RDBMS supports functions and encapsulation in restricted ways, where OODBMS offers full support to functions and encapsulation [14], in addition, RDBMS has no concept equivalent to inheritance. III.9. Security and authorization Although, with the magnitude of protecting sensitive data, Security and authorization have not been adequately addressed by vendors of OODBM, In contrast, many relational systems offer certifiable security levels. III.10.

Referential Integrity

Referential integrity applies to both RDBMS and OODBMS. RDBMS provides the following referential integrity actions for deletions and updating; Cascade which means deletion or updating of a record may cause deletion or updating of corresponding foreign-key records, No action, Set null or Set default in corresponding foreign-key records through deletion or updating process. Some OODBMSs support referential integrity in maintaining the dependencies between objects and avoiding dangling references. Some OODBMSs maintain the integrity of inverse members which look like cascade action on RDBMS [15]. III.11.

Performance

OODBMS is handling complex data by ten to a thousand times than an RDBMS, this high performance back to natural of OODBMS which convert the OIDs stored in an object to memory pointers when the object is loaded into memory. This feature does not exist in RDBMS. Also, even if OODBMS is not indexed, it may Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

execute queries by sequential scan. However, RDBMS outperforms OODBMS in creation, update, and deletes process [12]. III.12.

Query

A query language is not necessary for accessing data from OODBMS unlike RDBMS, it is still possible to use queries in an OODBMS but it faces the problem of weak query optimizers. The relational model has only a fixed set of operations in SQL, which are too restrictive to model the behaviors of object; also RDBMS faced the problem of handling recursive queries, where OODBMS faced the problem of lack of a formal mathematical foundation that leads to weaknesses in their query support [8].

IV.

Database Diagram

ISO 9001:2000 is one of Total Quality Management TQM approach. ISO 9001:2000 System consists of four sub-systems: planning system, documentation system, auditing system and correcting system. Planning system deals with choosing team and training processes. Documentation system deals with documentation processes such as writing the total quality management manual and decisions. Auditing system deals with auditing process. Corrective system deals with correcting processes. Documentation system works with two types of documents: The internal documents which generating inside the organization like topics in TQM manual, decisions and auditing documents. The external documents which generating from other organizations. Internal document issued by an employee and belong to a unit. Fig. 1 show the documentation system database represented in object model. Fig. 2 show the documentation system database represented in relation model. When applying the same task on the both models, each one handles the same task in a different way. For example, if the user wants to display all documents which belong to unit X and wrote by employee Y, system was built on OODB will search on all documents objects for document has the following data: Document.unit=& X and Document.employee=& Y. Where system was built on RDB, needs three steps to handle the same task. First, it will search on documents table for all documents have unit field equal to X regardless if it has employee field equal to Y or not. Second, it will search on documents table for all documents have employee field equal to Y regardless if it has unit field equal to X or not. Third, it will make intersect between both results. We can also note that part of ISO system in OODB looks simpler and easier to develop and maintenance.


1003


part of ISO 9001:2000 System. We can expect significant gains from the Object Oriented Database in comparison with the Relational Database. In particular, the Object Oriented Database can represent and handle the complex data in a high performance and can develop the systems faster with little code.

Person Fname : String Sname : String Tname : String Lname : String ID : Long Birthdate : Date Nationality : String Gender : String Address : String Social_status : String Children_num : Integer Religion : String Phone : String Mobile : String Email : String

References [1]

Document ID : Integer Date : Date belong to Title : String Type : String 0..* 1 Unit : Unit Content : BFILE

Unit member of ID : Integer Name : String 1..* 1 E-mail : String Phone : String 1 head of HoF : REF Employee 1

Add file()

Employee ID : Long JobDegree : Integer Job : String Salary : Currency Education _ level : String HigherDate : Date LeavingDate : Date Unit : REF Unit Account : Account Committee : REF Committee

[2] [3]

wrrote1by

[4]

1 Internald 0..* From : Employee

Externald From : String

Decision Meeting : REF Meeting

[5]

worked on

0..* Auditing_document Work_on : REF Employee

[6]

Topic Code : String Edition : Integer Status : String

[7] [8]

Fig. 1. Documentation system database in OODB externald ID Date tITLE Type 0..* Unit Content From

Import to 1

unit ID Name email 1 phone HOU 1

Generated by

[9]

Decision ID Date Title Type 0..* Unit Content Meeting 0..*

[10]

1

Done on

auditing 0..* document ID Date 0..* Done by Title Type Unit Content Work-on 0..* Work on From

1

1

[11] employee Fname Sname Issued by Tname Lname ID Birthdate Nationality Gender Address 1 Social_status Children_num Religion Phone Wrote by Mobile 1 Email IDE JobDegree Jobname Salary Education HigherDate LeavingDate Unit Username Password Power Committee

[12]

Belong to 0..*

0..*

topic ID Date Title Type Unit Content From Code Edition Status

[13]

[14] [15]

Authors’ information 1 Management Information Systems. Department King Abdulaziz University, P.O. Box 80201 Jeddah, 21589 Saudi Arabia. E-mail: [email protected]

Fig. 2. Documentation system database in RDB

V.

M. F. Worboys, H. M. Hearnshaw, D. J. Maguire, Object-Oriented Data Modeling for Spatial Databases, International Journal of Geographical Information Systems, Vol. 4, n. 4, pp. 369-383, 1999. S. Edlich, H. Horning, J. Paterson, R. Horning, The Definitive Guide to Db4o (Berkely: Apress, 2006). M. Alam, Migration from Relational Database into Object Oriented Database, Journal of Computer Science, Vol. 2, n. 10, pp. 781-784, 2006. N. Leavitt, Whatever Happened to Object-Oriented Databases, IEEE Computer Society DL, Vol. 33, n. 8, pp. 16-19, 2002. C. Chambers, D. Ungar, B. Chang, U. Hölzle, Parents are Shared Parts of Objects: Inheritance and Encapsulation in Self, Lisp And Symbolic Computation: An International Journal, Vol. 4, n. 3, pp. 207-222, 1991. S. Neelam, T. Benjamin, Testing Polymorphic Behavior, Journal of Object Technology, Vol. 1, n. 3, pp. 173-188, 2002. S. Herrmann, Gradual Encapsulation, Journal of Object Technology, Vol. 7, n. 9, pp. 47-68, 2008. T. M. Connolly, C. E. Begg, Database Solutions a Step-by-Step Approach to Building Databases (Second Edition, England: Pearson Education Limited, 2004). S. Luthra, Architecture In Object Oriented databases (2008), 16/3/2009.http://www.rimtengg.com/coit2007/proceedings/pdfs/6 3.pdf D. Obasanjo, An Exploration of Object Oriented Database Management Systems, (2001), 16/3/2009.www.25hoursaday.com/WhyArentYouUsingAnOOD BMS.html M. Stonebraker, L. A. Rowe, B. G. Lindsay, J. Gray, M. J. Carey, D. Beech, The Committee for Advanced DBMS Function: Third Generation Data Base System Manifesto , SIGMOD Conference (1990) S. Bagui, Achievements and Weaknesses of Object-Oriented Databases, Journal of Object Technology, Vol. 2, n. 4, pp. 29-41, 2003. S. Kazimierz, Impedance mismatch, (2008) 16/3/2009, http://www.ipipan.waw.pl/~subieta/SBA_SBQL/Topics/Impedanc eMismatch.html M. Blaha, The Dilemma of Encapsulation vs. Query Optimization, (2005), 1/3/2009, www.odms. org/experts.aspx#article1 M. Blaha, Referential Integrity Is Important For Databases, (2005), 16/3/2009, www.odms.org/ experts .aspx#article1

2 Information Systems, Department King Abdulaziz University, P.O. Box 42840Jeddah, 21589Saudi Arabia E-mail: [email protected]

Conclusion

Our objective in this paper is to study the differences between OODBMA and RDBMS by applying the both model on documentation system's database which is a



1004


Kamel Khoualdi was born in Algeria on 16 January 1966. He received his Ph.D. degree in computer science from Pierre & Marie Curie University (Paris 6), France, in 1994. He received his M.Sc. degree in Computer Science from Paris 6 University, France, in 1990. He received his B.Sc. in computer science from Batna University, Algeria, in 1988. His main research interests concern artificial intelligence, multiagent systems, knowledge-based systems, and e-learning. Thoria Alghamdi was born in Saudi Arabia on 12 August 1979. She received her M.Sc. degree in Business Administration from King Abdulaziz University, Jeddah, Saudi Arabia, in 2011. She received her B.Sc. degree in computer science from King Abdulaziz University, Jeddah, Saudi Arabia, in 2003. Her main research interests concern database systems, object-oriented analysis and design.



1005


A Hybrid Particle Swarm Algorithm for Job Shop Scheduling Problem Gongfa Li1,3, Yuesheng Gu2, Hegen Xiong1, Jianyi Kong1, Siqiang Xu1

Abstract – Production scheduling is a hotspot of manufacturing system and the core of the whole advanced manufacturing system. An effective scheduling method and optimization technology is the foundation and the key to realize advanced manufacturing and improve production efficiency. And algorithm research is one of the important content of the production scheduling problem. In recent years, various intelligent computation methods have been gradually introduced into the scheduling problem, such as genetic algorithm and simulated annealing algorithm, etc. The standard particle swarm optimization algorithm has a high computational complexity when used to solve the production of job-shop scheduling problem. The metropolis sampling criteria is introduced into the PSO algorithm. Other algorithms combined with particle swarm optimization algorithm, three kinds of fusion simulated annealing thoughts of hybrid particle swarm algorithm are constructed respectively. Comparing the results of hybrid PSO with the other algorithms in scheduling the job-shop benchmarking problem, the effectiveness and superiority of the hybrid Particle Swarm algorithm are verified. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Job Shop Scheduling, Particle Swarm Algorithm, Simulated Annealing Algorithm

I.

Introduction

At present, particle swarm algorithm has already got valid application in continuous space function optimized the question (number value optimized the question ), but there are still less research results in the field of production scheduling, and an important research direction of the production scheduling theory is an algorithm. So optimization method of particle swarm algorithm in production scheduling is very important. Particle swarm algorithm has the shortcoming of local optimization, hybrid algorithms is an effective solution .So it is the important research contents of particle swarm algorithm to study particle swarm algorithm combining with other algorithms and application. A discrete particle swarm optimization algorithm was presented for Job Shop scheduling problem. In the algorithm, a sequence—based code and update strategy for new positions were applied so as to make PSO more suitable for scheduling problems. Aiming at the shortcoming of premature and poor resulted from pure PSO, based on the complementary strengths of PSO and Variable Neighborhood Search (VNS) algorithm, four hybrid procedures were put forward. Numerical simulation demonstrated that within the framework of the newly designed hybrid algorithm, the NP—hard classic Job shop scheduling problem could be solved efficiently [1]. A hybrid quantum particle swarm optimization algorithm for the job shop scheduling problem is proposed and a coding method based on the order is designed.


1006

The simulation results show that the hybrid algorithm has good global convergence ability [2]. An easily implemented hybrid approach for the multi-objective flexible job-shop scheduling problem is developed. The results obtained from the computational study have shown that the proposed algorithm is a viable and effective approach for the multi-objective FJSP, especially for problems on a large scale [3]. A self-adaptive chaotic particle swarm algorithm for short term hydroelectric system scheduling is put forward [4]. A similar particle swarm optimization algorithm for permutation flowshop scheduling is put forward to minimize makespan [5]. Blending scheduling under uncertainty based on particle swarm optimization algorithm is put forward [6]. When standard particle swarm algorithm, which was just effectively applied to solving moderately simple production scheduling problems (such as problems of FT06 and LA01), was utilized to resolve complicated production scheduling problems (such as the problem FT20), it is extraordinarily difficult for the algorithm to converge at the global optimal solution. How to design the PSO algorithm suitable to solve complex the JSP problem becomes the problem which is researched in this paper. Here, the PSO algorithm combined with the simulated annealing algorithm, three kinds of hybrid PSO algorithms are designed so that solution of the complex JSP problem was achieved.


Gongfa Li, Yuesheng Gu, Hegen Xiong, Jianyi Kong, Siqiang Xu

II.

Design of PSO Algorithm Combining with Simulated Annealing Concept

A concrete problem optimized, how to make generated candidate solutions pervade the entire solution space is crucial to that the algorithm can effectively converging. The generation mode of initial solution depends on problem nature. This paper adopts the mode of randomly generating initial solution, which makes candidate solutions permeate the overall solution space. Aiming at the specific problem of JSP, the introduced hybrid PSO algorithm still selects the process-based coding mode and discrete-position evolutionary equation. The problem is discussed for the optimization objective of minimizing machining time. In order to surmount the PSO premature phenomena [7], PSO is combined with SA. At first, a superior population is obtained by making use of the fast research ability of PSO. And then a part of better individuals are optimized by utilizing the step ability of SA. Simulated annealing manipulation is to integrate the Metropolis sampling criteria into PSO algorithm, combine with the optimized objective function to compare adaptability magnitude between particles and self-particle optimal solution, and accept the solution larger than self-particle optimal solution according to the Metropolis sampling criteria so that it is guaranteed that PSO algorithm could effectively jump out of the local optimal solution. The PSO convergence rate is fast but its accuracy is inferior; SA is of powerful generality and easy to be realized. Nevertheless, its computational time is long and efficiency is lower. Moreover, SA becomes easily stuck in local optimization. Taking complementarities between PSO algorithm and SA algorithm in global search and local search into account, this paper combines with JSP characteristics, considers three kinds of modes which implement the conjunction of PSO and SA, and exerts sufficiently their advantages to construct hybrid discrete PSOSA schedule algorithm. II.1.

PSOSA Hybrid Algorithm-I

Because population optimal position is adopted in the location update formula, all particles have a tendency to fly into the population optimum position. If the population optimum position is located in local minimal solution, all particles tend to local minimal solution so that it induces search dispersity to become worse and makes global search ability of particles wane. Consequently, in order to boost the ability to avoid algorithm from be stuck in local minimal solution, we try to choose among many pbi a position labeled as gbi' ( t ) to substitute for gbi from the update formula. Therefore, the location update formula changes into equation (1): X i (t ) =

(

(

)

= ci ⊗ g ci ⊗ g w ⊗ f a,b ( X ( t ) ) , pbi ( t ) ,gbi' ( t )

)

(1)


Apparently, satisfactory performance pbi should be provided with superior selected probability. Utilizing the SA algorithm mechanism, we consider pbi which is worse than gbi as special solution. Accordingly, step probability of pbi comparing with gbi is calculated at the temperature t , namely, where f denotes objective function value. If the step probability is defined as pbi fitness value, pi of pbi replacing gbi is calculated in accordance with the following equation (2): pi = e

(

− f pbi − f gbi

)

t

∑e ( N

− f pbi − f pbi

)t

(2)

i =1

According to the substitution probability above, the roulette strategy is adopted to randomly ascertain which pbi becomes, which is good to overcome defects that PSO algorithm convergence rate is too fast and PSO algorithm cannot jump out of local optimal solution. Procedures of the roulette strategy selecting pbi are depicted as follows: Step 1: According to substitution probability pi of n particles we produce next generation p'i . Here: p1' = p1 , p'i = p'i −1 + pi , i=2,3,…n ; Step 2: Assign i =0;

Step 3: Generate a random number r among 0 ~ p'n ,

r = random ( 0 ,1) ;

Step 4: Compare r and p'i , If r < p1' , corresponding to p1 is selected as

gBi'

; If

p'i −1

pbi

< r < p'i ,

pbi corresponding to pi is selected as gBi' ,i=2,3,…n; Step 5: If i = n , end; else go to step 6; Step 6: Assign i = i + 1 , and go to step 3; To sum up, the solution flow of PSOSA hybrid algorithm-I is represented as following. Step 1: Initialization; Step 1.1: Generate N initial populations in accordance with the manner of generating initial solution and initialize inertia weight coefficient, acceleration constant, and coefficient of temperature drop; Step 1.2: Calculate adaptability magnitude of each particle in populations; Step 1.3: Assign optimal position pbi ( t ) of each

particle self and its objective value as the current position and its objective value. Assign population optimal position gbi ( t ) and its objective value as the optimal position and objective value in all pbi ( t ) ; Step 1.4: Ascertain initial temperature. Step 2: Update particles Step 2.1: Determine fitness value of each pbi ( t ) at the current temperature according to equation (2); Step 2.2: Adopt the roulette strategy to ascertain N


1007

Gongfa Li, Yuesheng Gu, Hegen Xiong, Jianyi Kong, Siqiang Xu

among all pbi t and update new position X i t of each particle in accordance with equation (1); Step 2.3: Calculate fitness value which the new position X i t of each particle corresponds to; Step 2.4: Update each particle pbi t and its fitness value as well as population gbi t and its fitness value; Step 2.5: Cooling manipulation. Step 3: If algorithm end condition is valid, gbi t and its objective value are output; else return to step 2. Adopting the procedure above to optimize the problem, we implement one selection of on each particle. Therefore, different particles maybe utilize different, which expands search dispersity to a certain extent. Meanwhile, there exists a certain contradiction between global dispersity search and local chemotaxis search. Hybrid algorithm adopts the SA algorithm mechanism to substitute various pbi for gbi . Contributing to surmounting premature convergence, it may give rise to the evolutionary process being extended. Difference of selected probability of all pbi is not obvious because of algorithm controlling temperature and higher temperature during early evolution. Thus, the algorithm emphasizes global dispersity search; as the temperature decreases, selected probability of pbi with good performance will increase and algorithm will emphasize local search of superior regions. Apparently, the temperature control strategy can automatically exert regulation on algorithm search behavior. The initial temperature t0 and cooling approach have a certain effect on the algorithm optimal performance. Here, we adopt experience formulae as following [8]: t0

f gbi ln 0.2

tk

1

tk

In order to boost PSO optimal performance, SA sampling process is executed on gbi after each iteration of the particle population and the attained result is defined as the new gbi of the PSO system. Thus, application of SA enhances search ability of algorithm for gbi so that probability of algorithm jumping out of local optimal solution is expanded. The flow chart of solution of PSOSA hybrid algorithm-II is shown as Fig. 2. Start

Select N initial solutions randomly and initialize w, c1, and c2. Ascertain iterative times, initial temperature, and cooling temperature coefficient l.

Calculate adaptability of various particles and get initial pbi(t) and gbi(t).

Yes

Iteration ends

Output optimal solution bi(t)

No Ascertain fitness value at the current temperature pbi(t)

Adopt roulette strategy to ascertain N gbi(t) from all pbi(t)

Update Xi(t) by utilizing discrete position equation and calculate adaptability

Update pbi(t) and fitness value

Update gbi(t) and fitness value

t=lt

Fig. 1. Solution Flow Chart of PSOSA-I Start Select N initial solutions randomly and initialize w, c1, and c2. Ascertain iterative times, initial temperature, and cooling temperature coefficient l.

(3) Calculate adaptability of various particles and get initial pbi(t) and gbi(t).

(4)

Iteration ends

Yes

Output optimal solution bi(t)

No Update Xi(t) by utilizing discrete position equation and calculate adaptability

where f gbi adaptation is value of the optimal particle in the initial population and is cooling rate.According to above-mentioned analysis of the solution process of PSOSA hybrid algorithm-I, the flow chart of solution of hybrid algorithm is demonstrated as Fig. 1.

Xi(t) ,i = 1, 2,3,…… ,n} , xi is the training sample,

IV.

SVMs are widely used in the field of text classification. In this paper, the fuzzy support vector machine based on convex hull is applied to test the classification testing set of text. The data that used in the experiment is UCI data set, programming language is Matlab, and the distance used in the experiment is Euclidean distance. We record the algorithm that proposed in this paper as HFSVM, and the fuzzy support vector machine based on class-center as FSVM. Having HFSVM, FSVM and SVM tested through the Breast, Heart and Pima data set in the UCI data set, these three data sets are shown in Table I.

Data Breast Heart Diabetes

yi is the class of the training samples, and si is the membership degree of the training samples. In order to obtain the optimal separating hyperplane, the optimization problem that shown in Eq. (6) need to be solved: n 1 2 min w + C Siξi 2 i =1

∑

subject to:

Experiment Result and Analysis

TABLE I DATA SET Number of Number of Class Attribute 9 13 8

Size of Sample

2 2 2

683 296 768

Classification precision is shown in Table II. TABLE II THE RESULTS WHILE KERNEL FUNCTION IS POLYNOMIAL

(6)

yi ( w ⋅ zi + b ) ≥ 1 − ξi , i = 1, 2 ,…… ,n

Data

SVM

FSVM

HSVM

Breast Heart Diabetes

94.74 74.32 69.03

96.49 74.32 77.6

97.31 83.24 82.54

ξi ≥ 0, i = 1, 2,……,n zi is the mapping relationship of training samples from original space to high-latitude space. Suppose the training sample set is {< yi ,xi ,si > ,i = 1, 2,3,…… ,n} , xi is the training sample, yi is the class of the training samples, the algorithm that proposed in this paper can be described as following: 1) Training the traditional SVM classifier, and divide the sample space for the first time to achieve the initial support vector and decision separating hyperplane wx + b = 0 ; 2) Calculating the center X of samples by using Euclidean distance and obtaining the radius R of hypersphere by Eq. (1). Then the sample set inside the hypersphere could be marked. 3) Constructing the convex hull of the sample set inside the hypersphere by using convex hull algorithm, and marking the sample points inside the convex hull. 4) Calculating the membership degree si of samples according to Eq. (4); 5) Introducing the membership degree of samples and rewriting the original training set as {< yi ,xi ,si > ,i = 1, 2,3,…… ,n} ;

6) Training fuzzy sample set, and constructing the optimal classification function and fuzzy classifier.


V.

Conclusion

Aiming at the problem that support vector machine is sensitive to noise, this paper introduces a fuzzy support vector machines based on dual split off sample space. After using the convex hull, the division of the sample space is more precise than the method that simply based on class center, the identification of noise is more accurate. And the experiments in UCI show that this method is valid in text classification.

Acknowledgements This work was supported by the Zhejiang provincial department of education scientific research projects (NO: Y201016682).

References [1] [2]

[3]

Lin CF, Wan SD, Fuzzy support vector machines, IEEE Trans. on Neural Networks, vol.13, n. 2, pp. 464-471, 2002. An Jinlong, Wang Zhengou, Ma Zhenping, Fuzzy Support Vector Machine Based on Density, Journal of Tianjin University, vol. 37, n. 6, pp. 544-548, 2004. Li Lei, Zhou Mengmeng,Lu Yanling, Fuzzy Support Vector Machine Based on Density with Dual Membership, Computer Technology and Development, vol. 19, n. 12, pp. 44-46, 2009.


1021

Hongyan Pan

[4]

[5]

[6]

[7]

Zhang Xiang, Xiao Xiaoling, Xu Guangyou. Fuzzy Support Vector Machine Based on Affinity Among Samples, Journal of Software, vol. 17, no. 5, pp. 951-958, 2005. Peng Xinjun, Wang Yifei, A Bi-Fuzzy Progressive Transductive Support Vector Machine Algorithm, Pattern Recognition and Artificial Intelligence, vol. 22, n. 4, pp. 560-566, 2009. Zhang Qiuyu, Jie Yang, Li Kai, Method of membership determination for fuzzy support vector machine, Journal of Lanzhou University of Technology, vol. 35, n. 4, pp. 89-93, 2009. Peng Xinjun, Wang Yifei. Total Margin ν-Support Vector Machine and Its Geometric Problem, Pattern Recognition and Artificial Intelligence, vol. 22, n. 1, pp. 8-16, 2009.

Authors’ information Department of Engineering, Zhejiang Business Technology Institute, Ningbo 315012, Ningbo, China.

Hongyan Pan, a banchlor of Zhejiang business technology institute, borned in 1979, graduated from Dalian technology university in 2005, majoring in information processing. Since 2005, more than ten papers have been published in some important journals. Two projects of hers was supported by Zhejiang provincial department of education scientific.



1022


Product Line Production Planning Model Based on Genetic Algorithm Guozhang Jiang1, Yuesheng Gu2, Jianyi Kong1, Gongfa Li1, Liangxi Xie1

Abstract – At present, iron and steel enterprise develops towards the direction with many procedure, many process, many variety and many specification, and reaches hundreds and thousands of product series and product mix, how to plan, organize and control steel production, production schedule of its product line are key issues. Mixed production plan model of product line can be summed up in a kind of network flow plan. According to the characteristics of network flow plan, the production schedule model of product network flow is established through describing digraph-connected graph of production procedure of iron and steel enterprise. Its goal function is the biggest profit of production of product line, restrain functions are the capacity limiting conditions, the balanced condition of the middle peak point, capacity restrain with supply and sell production and restrain with enterprise procedure process resources. Several key resource production procedure processes are chosen to calculate by using standard library function of the Matlab7.0 genetic algorithm toolbox to program. Penalty function is adopted in the course of getting solution. These parameters of scale of father population, crossover probability, mutation probability and penalty factor are combined and optimized. Results indicate that goal value reach convergence after finish 119 iterative operations. It accords with the actual conditions of this enterprise basically that the optimization solution to production plans of real iron and steel enterprise by using the algorithm. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Production Planning Model, Product Line, Genetic Algorithm, Iron and Steel Enterprise, Simulation

I.

Introduction

In recent years, Chinese iron and steel industry has obtained considerable development. With the restraining of surplus production, limited resource and insufficient transportation ability and, manufacturing enterprise needs dynamic, timely, ordered and integrated production strategy of production planning and scheduling urgently in order to realize improvement of the technological structure of iron and steel enterprise, advancement of product structure and increase of the production efficiency. Therefore, it’s necessary to pay high attention to production planning and scheduling of enterprises. A framework is presented to encompass this completely integrated system for using discrete event simulation as a modeling method. The system modeling framework addresses factors including customized configuration attributes and individual customer-preferred considerations for customized configurations. The framework is intended to aid decision-making concerning cost and schedule impacts associated with customization options chosen throughout the supply chain.


1023

A real-world example drawn from aerospace is included to demonstrate and validate the operational capability of the proposed framework [1]. The production line of companies which produce high-quality rolled aluminium lithographic strips is studied. An efficient representation for such production processes is provided and subsequently used for an extensive analysis and performance evaluation through appropriate metrics. In particular, the work addresses the implementation of an overall model in a simulation environment, capable of integrating the various aspects of the specific production management processes. The model was successfully validated using actual production data, and it was found that it is suitable for the modeling, analysis and performance evaluation of the complex aluminium coils production process. With the aid of the model, various scenarios were investigated via extensive simulation runs, such as installing additional machine centers and reducing pre-set times the products spend in intermediate storage areas [2]. Managing uncertainty is a main challenge within supply chain management. Therefore, it is expected that those supply chain planning methods which do not include uncertainty obtain inferior results if compared with models that formalize it implicitly. This article presents a review of the literature Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Guozhang Jiang, Yuesheng Gu, Jianyi Kong, Gongfa Li, Liangxi Xie

related to supply chain planning methods under uncertainty. The main objective is to provide the reader with a starting point for modeling supply chain under uncertainty applying quantitative approaches. We have defined taxonomy to classify models from 103 bibliographic references dated 1988–2007. Finally, some conclusions about the works analyzed have been drawn and future lines of research have been identified [3].Thus, production planning in manufacturing involves in most cases the synchronization with the downstream demand and thereby has a strong impact in warehouses of both manufacturers and other participants of supply chains [4]. A more detailed task in manufacturing is production scheduling where managers in the context of the optimal production planning must couple individual products with individual productive resources in the shortest times. Wu et al. [5] suggest that scheduling can be a cumbersome task especially in cases when last minute changes are imposed by machine breakdowns, new high-priority orders arrival and the occurrence of other disruptions. The serialization and diversification of product is the developing trend of enterprise production. The product line is a product family with multi-varieties, multi-steel-categories, multi-specification of iron and steel product. It is useful to enhance research ability and adaptive ability for the market and increase the enterprise benefits if plan and management have been done to product line in steel enterprises. Fig. 1 is the product line of a steel enterprise and it shows total product manufacturing in this enterprise. In the plan and management of an enterprise product line, the product line combining production planning model is the key point of product line combining strategy. When building a model, the final product (finished product in sale) is usually chosen as planning object.

100. If the availability of complex production line are considered too, it can be forecasted that the number of product planning objects can reach to 103~106 [6].

II.

Proposed Network Flow Model of Product Combining Product Plan

Based on graph theory, network flow planning is actually a linear programming problem, but it specially deals with optimizing problem with the characteristic of network using graph theory, so it is convenient to deal with complex constraint problem with high dimension [7], [8], and network flow is much faster to find the solution than normal linear programming because it built a series of algorithms focusing on network. Thinking about the available process paths from raw materials to finished products, there are different equipments, raw materials or semi-manufactured products to couple the material requirements planning with task accreditation planning, therefore, some model algorithms will not work or be hard to find the solution. At the same time, network flow planning has additively and infinite decomposability of network and its algorithms are easy to understand and simply to find the solution. Hence, network flow optimizing method is adopted to analyze product combining planning. In the problem of network flow planning, the start point is called as Source Point represented by S , while the end point is called as Sink Point represented by T , and the normal node is apex or intermediate point, which uses i to indicate, and the point next to point i is point j . In the course of modelling, there is some hypothesis as following: (1)Provided that infinite materials, energy, information and manpower resource expect to export in the source point, while there is an infinite storage to store these materials, energy, and information and manpower resource. (2)Provided that intermediate points with no ability to store materials, energy and information, just play a role of transportation. (3)The maximum transportation amount is prescribed as the capability of the path from point i to point j , which is represented by C ( i, j ) . (4) X ( i, j ) is defined as factual transportation flow from point i to point j . Network flow model is shown in Fig. 2.

Fig. 1. Product line of an iron and steel enterprise

The iron and steel products usually have the information about variety, specification and so on, and the amount of information combining is huge. Taking Baosteel as an example, if the varieties sold every year are more than 1,000, and the combining amount of the sub-specification in the same variety will be more than


Fig. 2. Network flow model


1024


The production process of iron and steel enterprise can be described as directed graph G = (V , A,D, p,U ,c ) [9], [10], where, V represents the set of materials, work-in-process (WIP) or finished products, and arcs A is the set of production process procedures. The subset of Apex set S ⊂ V is supporting or selling materials, which is intituled as Source Points, nonnegative number Di indicates the supporting ability or selling ability of apex , i ∈ S , and nonnegative number pi indicates the bargain price of apex i ∈ S . Nonnegative number U ij is production ability of flow procedure, and nonnegative number cij indicates production cost per number of units. Given the flow of either arc ( i, j ) ∈ A in G is xij , the

{ }

colony x = xij is called as a network flow of network G . The network flow production planning model is built as follows: ⎛ ⎞ MinE ( f ) = ∑ pi ⋅ ⎜⎜ ∑ xli − ∑ xim ⎟⎟ − ∑ cij ⋅ xij (1) i∈S m ⎝ l ⎠ ( i , j )∈A

s.t.: 0 ≤ xij ≤ U ij

(2)

∑ xli = ∑ aim xim

(3)

l

m

∑ xli − ∑ xim l

∑

( i, j )∈A

x01,02 ≤ a1

(6)

Wire, bar, section steel flow procedures constraints: x1112 , + x1113 , + x1114 , + x1115 , ≤ a2

(7)

Media plate flow procedure constraints:

x09 ,10 ≤ a3

(8)

Hot rolling flow procedure constraints: x03,05 + x03.04 ≤ a4

(9)

where, the value of a1 ,a2 ,a3 ,a4 is 500, 300, 100, 200 respectively.

III. Design and Solution of the Genetic Algorithm (1) Chromosome Encoding Design There are 15 variables in the model and they are all real number, so dimension of the model is high. For some high dimension continuous function optimizing problems requiring high precision, binary encoding can be used except some disadvantages, and float encoding can also be used.

(2) Initial Population ≤ Di , i ∈ S

(4)

m

gijk ⋅ xij ≤ bk ; k = 1, 2,… ,K

(5)

where, nonnegative number aim is plus parameter of every arc, when the flow is constant, aim =1. Nonnegative number gijk is the amount of consuming resource k in production per unit j through process procedure ( i, j ) , and nonnegative number bk is the amount of resource k . The goal function is described as equation (1), which means the maximum profits of production plan, formula (2) indicates capability constraint condition, formula (3) indicates either intermediate apex balancing condition, formula (4) constrain either supporting or selling ability, and formula (5) indicates any resource constraint of enterprise. In fact, the author chooses the resource constraints of some key flow procedures, such as steelmaking, wire, bar, section steel flow procedures, media plate flow procedure and hot rolling flow procedure described as formula (6)~(9). The maximum of profits production plan is calculated using above algorithm: Steelmaking flow procedure constraints:


Based on the genetic algorithm toolbox of Matlab 7.0, function gaoptimset ('PopulationSize', L) is used to create initial population with L chromosomes at random.

(3) Treatment of Fitness Function In the course of settlement about model constraints, adopting penalty function to reduce the fitness ability of individuals that don’t satisfy constraints, then their reproducing can be controlled. How to build a fitness penalty function is the key problem. Penalty function form is as follows: penalty f ( x ) = ∑ j =1 Rij f j2 ( x ) m

(10)

‘ i ’ different constraint-against level to the j th constraint at first is built, different penalty coefficient Rij is determined focusing on every level. The higher the constraint-against level is, the greater the value of Rij is. The penalty function used in program can be described as formula (11): penaltyf ( x ) = ⎧⎪rho ∗ sum ( abs ( ceq ) ) Dissatisfy ec =⎨ ⎪⎩mu ∗ sum ( c ( c < 0 ) ) ^2 Dissatisfy ic

(11)


1025


In formula (11), rho is equation penalty coefficient, while mu is inequation penalty coefficient; ec is the equation constraint, and ic is the in equation constraint. the main computing codes are as following: function f = fcxyzh(x，rho，mu) rho =11; % RHO parameter is used for equality constraints mu =7; % MU parameter is used for inequality constraints of type c(x) >=0 % x is the design variable and rho and mu are the penalty factors % Evaluate the goal first f=-(3620*x(1)+3630*x(2)+3410*x(3)+3400*x(4)+45 00*x(5)-740*x(6)-20*(x(7)+x(8)+x(9)+x(10)+x(11))); [ceq ， c] = mycon(x) ； % Evaluate the nonlinear equality and inequality constraint(s) f = f + rho*sum (abs (ceq)); % Add equality penalty f = f + mu*sum(c(c 0 Ws, stem implant width Hs, stem implant height M, Magnification factor 30/1, 24/1, 18/1 Pr, Pixel ratio between computer display and digital X-ray

Ws Fig. 15. Digital CPT implant



1075

A. Shapi’i, R. Sulaiman

Wna = Wa × M × Pr Hna = Ha × M × Pr Wna, Hna, Wa, Ha, M, Pr > 0 Wa, acetabular implant width Ha, acetabular implant height M, Magnification factor 30/1, 24/1, 18/1 Pr, Pixel ratio between computer display and digital X-ray By using this algorithm, the digital implant can be scaled accurately. Output for the two algorithms above will be discussed in the subsequent section.

IV.

Results and Discussions

The output of the implant magnification algorithm is shown in Fig. 17 (default size). In Fig. 17, the X-ray image size is 16.7% of the actual size. Fig. 18 shows the size of medical images zooming to the level of 100%. Here, it can be seen that when the size of the digital medical images is enlarging, the digital implant also grows on scale as determined by the magnification algorithm.Therefore, the use of digital implants with 600 dpi or higher is appropriate. As we can see in Fig. 18, the implant image is not distorted even when the medical images zoom to the level of 100%. Based on Fig. 17 and Fig. 18, the scaling method developed in this research is suitable to scale the implants with medical images and computer display. The technique developed enables accurate implant scaling and is suitable to be used with digital medical images because of the ability of the algorithm to read the resolution information and X-ray image pixel density. In order to test the accuracy of the digital implant an experiment was conducted with assistance of an experienced surgeon through the conventional method to determine the stem implant size. The results through the conventional approach were compared to the results produced by our digital method.

Fig. 18. X-ray image zooming at level 100%

The testing recorded the implant size to be used and the time taken by both methods. For both implant components twenty randomly selected X-rays of unidentified patients were used for templating both techniques. The difference between the two sizes was calculated and summary is shown in Table II and Fig. 19. It is evident that the new digital technique yields very close results to those obtained through the conventional method in all twenty studies. The difference, if any, is also within the error of clinically acceptable range (±1 for CPT stem and ±2 mm for acetabular implant) obtained through the conventional templating method. In addition, the study also demonstrated that the average time taken for implant templating in THA pre-operative planning using digital implant was much less than when using the conventional method. TABLE II DIGITAL VS CONVENTIONAL Stem CPT Similar Size

Acetabular

11

55%

Size ± 1

8

45%

Size ± 2

1

5%

Similar Size

12

60%

Size ± 1

8

40%

Size ± 2

0

0%

100 90 80 70 60 50 40 30 20 10 0

CPT Stem Acetabular

Similar Size ±1 / Size ±2 / Size ±2 mm ±4 mm Fig. 17. X-ray image zooming at the level of 50%

Fig. 19. Effectiveness of new digital technique



1076


V.

Conclusion

References

AutoCAD is highly skilled software for performing drawing works. AutoCAD usage is extremely useful in designing digital CPT implants for hip joints. The research shows that the AutoCAD 2010 software has a high potential for producing a detailed and complex product design. At present, the AutoCAD usage is not only limited to the field of manufacturing and architecture but can also be used in the medical field. The production of CPT implants allows it to be used in a digital environment [14]. The digital implant provides several advantages for THA surgery. Compared to the conventional method in which the surgeon uses a template manually and places it on the patient’s X-ray, the use of digital implants not only saves time but can also reduce the errors due to consistency difference when making adjustments to a patients implant size [15]. In addition, by using digital implants, it can be manipulated easily for things such as doing rotation and scaling. Scaling is one of the geometric transformations which can be used to shrink or enlarge an object according to the specifications or the scale ratio specified by the user. in this study the implant magnification technique is used to scale the digital total hip implant with x-ray medical images. The fact that x-ray images scale conversion to the actual size of the bone is the basis for the provision of a digital templating process [16]. The use of x-ray images, which are converted to jpeg format may cause an error with regards to scaling implants. The provision of digital implants with high dpi value is highly recommended. the higher the value of dpi, the higher the quality of the resulting digital implants. Based on the output generated from the implant magnification techniques in this research, it can be said that the technique is capable of accurate implant scaling. implant size will reduce or enlarge according to x-ray image size. With greater emphasis on the resolution and pixel density images, this technique enables a reduction of errors during preoperative evaluation. In addition, the use of quality X-ray images can also have a strong impact on the production of scaling techniques with high accuracy. The digital method, if being used properly, will enable us to help surgeons to make decisions accurately and effectively.

[1]

[2]

[3] [4]

[5]

[6]

[7]

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

J.A. Davila, M.J. Kransdof, and G.P. Duffy, Surgical Planning of Total Hip Arthroplasty: Accuracy of computer assisted EndoMap software in predicting component size, Journal of Skeletal Radiol, Vol. 35, pp 390-393, 2006. L.W. Carter, D.O. Stovall, and T.R. Young, Determination of accuracy of preoperative templating of noncemented femoral prostheses, Journal of Arthroplasty, Vol. 10, n.1, pp. 507513.1995. Y. Siti Fairuz, Pengautomasian Templat Penggantian Sendi Lutut, Masters thesis, Universiti Kebangsaan Malaysia, Malaysia, 2009. J.A. Hendrikus, Crooijman, M.R.P. Arman, C.V. Pul, and J.B.A Mourik, A New Digital Preoperative Planning Method for Total Hip Arthroplasties, Journal of Clin Orthop, Vol. 467, pp. 909916, 2008. W. Christian, Q. Henning, J. Xu, H. Hansjoerg, V.K. Marius, and S. Guido, Digital Templating in Total Hip Arthroplasty with the Mayo stem, Journal Orthopaedic and Trauma, Vol. 3, pp. 10231029, 2008. J. Arora, The Role of Pre-operative Templating in Primary Total knee Replacement, Knee Surgery Sports Traumatol Arthosc, Springer, pp.187-189, 2004. R.N.J. Graham, R.W Perriss, A.F. Scarsbrook, DICOM Demystified: A Review of Digital file formats and their use in radiological practice, Clinical Radiology, pp.1133-1140, 2005. B.F. Kavanagh, Femoral fractures associated with total hip arthroplasty. Complications of Total Hip Arthroplasty, Orthopaedic Clinics of Americ, Vol. 23, pp. 249, 2005. C.S. Krishnamoorty, Computer Aided Design: Software and Analytical Tools. 2nd Edition (United States: Alpha Science International Ltd. 2004). M. Klein, Using imaging data in Making Orthopedic Diagnoses. In DC Chhieng, and GP Siegal, Advances in Experimental Medicine and Biology, Vol. 44, pp. 104-111, 2005. Y. Ilchman, C. Eingartner, K. Heger and K. Weise, Femoral Subsidence Assessment After Hip Replacement Upsala J Med Sci, Vol. 111, pp. 361-369, 2006. Y. Jun, K. Choi, Design of Patient specific hip implants based on the 3D geometry of the human femur, Journal of Advances in Engineering Software, Vol. 41, pp. 537-547, 2010. J.R. Verdonschot, P. Horn , R.L. Oijen, and Diercks, Digital analogue Versus Preoperative Planning of Total Hip Arthroplasties, The Journal of Arthroplasty, Vol. 22, pp . 866870, 2007. S. Azrulhizam, S. Riza, H. Khatim, S. P. Anton, M.K. Yazid and M.H. Hazlah, “Design of Digital Implant for Pre-Operative Planning in Total Hip Replacement,” in Proceedings of International Conference of Electrical Engineering and Informatics (ICEEI 2011) (2011). A.G. Valle, F. Comba, N. Taveras, E.A. Salvati, E.-A, The Utility and Precision of Analogue and Digital Preoperative Planning for Total Hip Arthroplasty, International Orthopaedics (SICOT), Vol. 32, pp. 289-294, 2007. A. Shapi’i, R. Sulaiman, M.K. Hasan and A.Y.M. Kassim, Scaling Technique for Digital Implant in Medical Images Using Pixel Density Algorithm, European Journal of Scientific Research, Vol. 47, n. 1, pp. 24-32, 2010.

Authors’ information

Acknowledgements This research project was conducted in collaboration with Dr Abdul Yazid Mohd Kassim and Dr Nor Hazla Haflah from the Department of Orthopaedics and Traumatology and Dr Hamzaini Abd Hamid from the Department of Radiology, Medical Centre of University Kebangsaan Malaysia. This research was also funded by the University Grants UKM-OUP-ICT-35-179/2010 and UKM-GUP-TMK-07-01-035.


1

School of Information Technology, Universiti Kebangsaan Malaysia. 2 Institute of Visual Informatics, Universiti Kebangsaan Malaysia.


1077


Mr Azrulhizam Shapi’i is currently pursuing Ph.D in Industrial Computing at School of Information Technology, Universiti Kebangsaan Malaysia. He is working as Lecturer in the School of Information Technology , Faculty of Information Science and Technology, University Kebangsaan Malaysia. His research areas of interest include Computer Aided Design, Medical Imaging, Computer Aided Medical System and Programming. Dr. Riza Sulaiman did his Msc from University of Portmouth, UK and Ph.D from University of Canterbury, New Zealand. His specializations include Computer Aided Design (CAD), Medical Imaging and Robots Simulation. He is working as Associate Professor in the Institute of Visual Informatics, Universiti Kebangsaan Malaysia.



1078


Improvement of Wood Ultrasonic CT Images by Using Time of Flight Data Normalization Honghui Fan1, Hongjin Zhu1, Guangping Zhu1, Xiaojie Liu2

Abstract – The maximum likelihood expectation maximization (ML-EM) algorithm was applied to ultrasonic time of flight (TOF) computed tomography (CT) for wooden pillars. The sound velocity changes with the direction of the ultrasonic propagation path, therefore, when the image was reconstructed by TOF data based on ML-EM method, the image had many artifacts. For the purpose of reducing the artifacts, we proposed a "gap average velocity" method in imaging process. Transmission paths of an ultrasonic wave through a cross-section of wood were corrected after TOF data normalization. The feasibility of TOF data normalization and TOF data interpolation were examined in detail by using wooden phantoms. The artifacts evidently disappeared and high quality reconstructed images of wooded pillars were improved by the proposed technique. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Ultrasonic Computed Tomography, Maximum Likelihood Expectation Maximization, Time of Flight, Normalization

I.

Introduction

Computed tomography (CT) is an emerging technique for non-destructive evaluation of wood. The objective of tomography is to provide visualization by cross section of the structure interior. There are some reports for the nondestructive inspection of wooden pillars using ultrasonic [1], [2]. Tomikawa et al. applied ultrasonic time of flight (TOF) computed tomography (CT) to wooden pillars [3]. Yanagida et al. reported that the ring shaped artifact in the CT image of wood was caused by orthotropic acoustic property of wood. The filtered back projection (FBP) algorithm was used for reconstruction method in their report [4], [5]. However, the reconstructed image of FBP has a problem of the emergence of negative pixel values [6], [7]. Maximum likelihood-expectation maximization (ML-EM) is an image reconstruction method for ultrasonic TOF CT. The ML-EM algorithm has been widespread applied in the medical field. This algorithm gives positive pixel values in addition to the fact that absorption, scattering, and resolution can be corrected in the ML-EM process [8], [9]. So ML-EM algorithm use iterative algorithms can generally give better images compared to the traditional FBP methods. However, the perceived quality of one CT image depends on the quantity of TOF data. And some artifacts appeared in the reconstructed images. In this paper, we proposed an approach for wooden inspection by using ultrasonic CT based on ML-EM. We obtained 306 TOF data in our experimental system, and all ultrasonic TOF data grouped with a gap.


1079

The sound velocity in wood is different with the direction of the ultrasonic propagation path [10]. We used "gap average velocity" method in imaging process to reduce the influence of the anisotropic acoustic property. The effects of image quality after TOF data normalization and interpolation were examined in detail using wood phantoms.

II.

Principle

We described the maximum a posteriori (MAP) algorithm in Fig. 1 and Fig. 2. In our imaging system, i is the index of a measurement pass, j is the index of a pixel, λ j is the slowness at the j-the pixel, yi is the TOF data of the i-th pathway, and Cij is the contribution rate of the j-th pixel to the TOF value obtained at the i-th pass. The ML-EM algorithm iteratively corrects the estimated slowness λ k +1 from λ k using the algorithm given by: λ(j

k +1)

=

k λ(j ) n

(k )

∑ Cij i =1

k yi Cij( )

n

∑ i =1

m

(k ) (k )

∑ Cil l =1

(1)

λl

where k is the iteration number. The TOF contribution of j-th pixel to the i-th TOF value is shown as: xij = Cij λ j

(2)


Honghui Fan, Hongjin Zhu, Guangping Zhu, Xiaojie Liu

In Fig. 2, as an example the projection of six pixels was assumed, and the pixel j=4 was requested. matrix of reconstructed image j（ number of pixel）

1 2 3

64

12 8

λj

j=4096

( θ c = 10° ) on the circumference of the test sample and labeled 0 to 35 (Fig. 3). Depending on the interval of transducers for transmission and reception, all sound velocities were grouped with a gap angle θ . When 36 measurement points are used, the number of groups is 9. The method of measurement of TOF data was the same as that explained by Yanagida et al [5]. The TOF data interpolation explained by Fan et.al [10], the TOF data increased to 2556 [=（transmitter 36 × receiver 71）/ exchange 2] after 5° interpolation.

Cij

θ c = 10 °

measuring point(0~35)

65

detector

1 2 3

64

xij

Transducer for reception Transducer for sending

Fig. 1. Notation and coordinate system for ML-EM reconstruction

The interval of transducer for transmitting and receiving is expanded.

λ 6 xi6 = Ci 6λ6 xi 5 =Ci5λ5 xi4 = Ci4λ4 xi 3 = Ci3λ3 xi2 = Ci2λ2 xi1 =Ci1λ1

j= λ 4 4

The interval of the transducer for transmitting and receiving is made constant; the data of one lap is collected.

P rojection

yi = Ci1λ1 +Ci 2λ2 +"+Ci 6λ6

xi 4 =Ci4λ4 i

i

Ratio of xijand yi

Fig. 3. Measurement procedure

yi Ci 4 λ4 xi 4 = Ci 1λ1 + Ci 2 λ2 +"+ Ci 6 λ6 general formula

xij =

yiCijλj m

∑Cij′λj′ j′ =1

Fig. 2. Procedure of MAP estimation

Detection probability Cij is described at the area rate that the TOF data of i-th expects pixel j. The propagation pathway is scanned at detailed intervals. When all the areas are covered, it is shown =1.0. To pixel j near the point in the pathway, Cij was calculated using:

(

Cij = ∑ 1 − x − x j

)(1 − y − y ) × s j

(3)

where, (x, y) is the position coordinate of the scanned position and (xj, yj) is the center of the j-th pixel.

The sound velocity in woods changes with the direction of the ultrasonic propagation path. Since the spatial distribution of sound velocity was reconstructed by ultrasonic CT, some ring artifacts appeared in the image because of the anisotropic acoustic property of wood. Yanagida et al. performed the "0 or 1 imaging method" to reduce the artifacts of reconstructed images based on FBP method. When two different sizes of defects are placed on the same circumference, the larger defect was emphasized, but the smaller defect did not appear in the reconstructed image [5]. To reduce the artifacts and accurate reconstruct defects image, the "gap average velocity" method was performed in our system. According our measurement method, all ultrasonic TOF data grouped with a gap angle θ c of the transmitter and the receiver (measuring angles from 20 to 180 with 20 intervals). The sound velocity vi (

gap )

must be almost the same for one same gap

pathways if the sample is normal (no defect). If vi (

gap )

( ) was considerably slower than the vave (average velocity of the gap), some defect should be on the i-th pathway. The threshold sound velocity was determined using: gap

III. Experimental Methods Measuring angle θ is the angle between two lines from the center of specimen to each measurement point. The thirty-six measurement points were placed 10° apart


gap ( gap ) vth( ) = δ vave

(4)


1080


1 35 ( gap ) ( gap ) vave = vi 36 i = 0

∑

(5)

artificial defects were 18 mm, 22 mm, 33 mm, and 40 mm. The diameter of the inspection object was 220 mm.

Here, δ was decided empirically to be 92%. Because δ =0.92 was obtained using the five specimens of one kind of wood (black locust), δ =0.92 might be adoptable for a few specimens. A common δ should be determined using much kind of woods. If vi (

gap )

path passed a clear area. On the other hand, if vi (

gap )

gap )

(b)

gap = vth( ) was used for the clear pathway instead of

the TOF value, and p 2i ( defect. That is: pi(

gap )

gap )

= vi(

gap )

was used to for the

( ) ≥ vth( ) ⎧ p1 for vi =⎨ i ⎩ p 2i for vi( gap ) > vth( gap ) gap

gap

(6)

Nine pi( ) groups were obtained after TOF data normalization for one CT image. As the influence of the anisotropic acoustic property was reduced, the artifact level of the reconstructed image decreased. gap

IV.

Figs. 4. Inspection object

was

gap slower than vth( ) , the path contained a defect area. So

p1i (

(a)

was faster than vth( gap ) , we determined that the

Results and Discussion

Ultrasonic frequency was 68 kHz in actual measurement, and the aperture dimension of the transducer was about 15mm, the wavelength in the wood is about 20mm, so the transducer was considered as omnidirectional. Five pieces of hardwood were prepared in our system. Four pieces of hardwood (Robinia psedoacacia L.) with one artificial defect in each piece were prepared as test specimens [Fig. 4(a)]. The center of the artificial defects was set 50 mm away from the center of the wooden pillars. The diameters of the

The number of TOF data after 5° interpolation was 2556. Using interpolated data, better quality images were reconstructed. The images reconstructed by using the ML-EM (5º) method compared with those reconstructed ML-EM (20º) method in Fig. 5. The images reconstructed by using the ML-EM (20º) method were used for only 306 measured TOF data. The defects were not clearly observed by using only 306 measurement TOF data. By using the TOF data interpolation of 5º, the extent of defect could be recognized in the CT image. The extent of defect could be recognized in the CT image using interpolated data in ML-EM algorithm. But, it was possible that some artifacts were recognized in the images. To reduce the artifacts, the "gap average velocity" method was used in our imaging process. The reconstructed images based on ML-EM algorithm with "gap average velocity" method are shown in Fig. 6. By using the method of "gap average velocity" based on ML-EM, we could clearly reconstruct the defect with 2556 TOF, and the artifacts disappeared [Fig. 6]. Further, to confirm the "gap average velocity" method, one piece of hardwood with two artificial defects was prepared as an additional test specimen [Fig. 4(b)]. The diameters of the defects were 20 mm and 33 mm, and the defects were placed at 90° intervals. The reconstructed images based on ML-EM method are shown in Fig. 7. From the reconstructed images, the two defects of different size could be easily recognized after 5º interpolation and TOF data normalization.

Fig. 5. Reconstructed images of one defect



1081


Fig. 6. Reconstructed images with "gap average velocity" method

Fig. 7. Reconstructed images of two defects in one specimen

V.

Conclusion

When the ML-EM method was used for the ultrasonic CT of wooden pillars, the artifacts caused by the anisotropic property of wood disappeared when the data normalization method was used, and high quality images were obtained. By using this technique, the accuracy of defect detection in wood can be significantly improved.

Acknowledgements The authors are very thankful to Tamura-Yanagida LAB from Yamagata University of Japan for providing experimental data.

References [1]

[2] [3]

[4]

[5]

[6]

K. Kim, S. Lee, J. Lee, Cross-sectional image reconstruction of wooden member by considering variation of wave velocities, Journal Korean wood Science Technology, vol. 35, n. 5, pp. 16-23, 2007. V. Bucur, Ultrasonic techniques for nondestructive testing of standing trees, Ultrasonic, vol. 43, n. 4, pp. 237-239, 2005. Y. Tomikawa, Y. Iwase, K. Arita,et al., Nondestructive inspection of a wooden pole using ultrasonic computed tomography, IEEE Transactions, Ultrasonic Ferroelectrics. Frequency Control, vol. 33, n. 4, pp.354-358, 1986. Y. Tamura, K. Adachi, Y. Yanagiya, M. Makino, and K. Shioya, Ultrasonic timeof-flight computed tomography for investigation of wooden piallars: Image reconstruction from incomplete time-of-fliht profiles, Japanese Journal of Applied Physics, vol. 36, pp. 3278-3280, 1997. H. Yanagida, Y. Tamura, K. Kim, and J. Lee, Development of Ultrasonic Time-of-Flight Computed Tomography for Hard Wood with Anisotropic Acoustic Property, Japanese Journal of Applied Physics, vol. 46, pp. 5321-5325, 2007. P. Razifar, M. Sandstrom, Harald Schnieder, Bengt Langstrom, Enn Maripuu, Ewert Bengtsson, Mats Bergstrom, Nosi correlation in PET, CT, SPECT and PET/CT data evaluated using autocorrelation function: a phantom study on data, reconstructed using FBP and OSEM, BMC Medical Imaging, vol. 5, 2005.


Available at: http://www.biomedcentral.com/1471-2342-5-5. Accessed June 7, 2007. [7] M. Kiguchi, K. Taniguchi, Y. Akiyama, T. Furukawa and T. Kushima, Evaluation of the number of sampling step angles in SPECT Images: comparison between FBP and OS-EM reconstruction algorithms, Jappan Journal of Radiological Technology, vol. 60, n. 7, PP. 1009-1016, 2004. [8] K. Matsumoto, T. Fujita, K. Oogari, Evaluation of Median Filtering after Reconstruction with Maximum Likelihood Expectation Maximization(ML-EM) by Real Space and Frequency Space, Jappan Journal of Radiological Technology, vol. 58, n. 5, pp. 670-678, 2002. [9] Y. Bazi, L. Bruzzone, F. Melgani, Image thresholding based on the EM algorithm and the generalized Gaussian distribution, Pattern Recognition, vol. 40, n. 2, pp. 619-634, 2007. [10] HongHui Fan, H. Yanagida, ShuQiang Guo, Y. Tamura and T. Takahashi, Image Quality Improvement of Ultrasonic Computed Tomography on the Basis of Maximum Likelihood - Expectation Maximization Algorithm Considering Anisotropic Acoustic Property and Time-of-Flight Interpolation, Journal of Applied Physics, vol. 49, pp. 07HC12, 2010.


College of Computer Engineering, Jiangsu Teachers University of Technology, Changzhou 213001, China. 2

Department of Electrical Engineering, Jiangsu Teachers University of Technology, Changzhou 213001, China. Honghui Fan was born in1980. He has a M.Sc. And Ph.D. from the Yamgata University of Japan in 2008 and 2011, respectively. In 2011 he joined Jiangsu Teachers University of Technology as a lecturer. His research interests include ultrasound imaging, image restoration and signal processing in biomedical engineering.


1082


Hongjin Zhu was born in 1981. She has a M.Sc. And Ph.D. from the Yamgata University of Japan in 2007 and 2010, respectively. She was employed as a special researcher in the Department of Engineering, Yamagata University of Japan in 2010. In 2011 she joined Jiangsu Teachers University of Technology. Her research interests include image processing, computer vision, pattern recognition and evolutionary computation. Guangping Zhu was born in 1965. She has a M.Sc. From the Sichuan University of China in 1999. She is a member of Jiangsu Computer Federation. And she is an Associate Professor of computer science at Jiangsu Teachers University of Technology. Her research interests include data mining and computational, fuzzy information processing. Xiaojie Liu was born in 1978. He has the M.S. degree and Ph.D. from the Jilin University, China, in 2006 and 2009, respectively. He has held lecturer at Jiangsu Teachers University of Technology, China, since 2009. His research interests include intelligent information processing, embedded visual system and UAVs.



1083

International Review on

Computers and Software (IRECOS)

PART

B

Contents (continued from Part A) Outline-Based Text Image by Guorong Xiao, Xuemiao Xu

1084

Integrated Price Forecast Based on Dichotomy Backfilling and Disturbance Factor Algorithm by Quanyin Zhu, Suqun Cao, Pei Zhou, Yunyang Yan, Hong Zhou

1089

A Target Location Method Based on Binocular Stereo Vision for Eggplant Picking Robot by Jian Song

1094

Credit Card Risk Detection Based on Chaos Theory and Cloud Model by Yanli Zhu, Yuesheng Gu, Shiyong Li, Shunping Wang

1099

Buyer-Seller Watermarking Protocol without Trust Third Party by Lv Bin, Fei Long

1104

Ontology-Based Oracle Bone Inscriptions Machine Translation by Jing Xiong, Lei Guo, Yongge Liu, Qingsheng Li

1108

A Method for Mining Association Rules Based on Cloud Computing by Fei Long, Yufeng Zhang, Lv Bin

1112

Software Quality Assurance Based on Java Modeling Language and Database by Shukun Liu, Xiaohua Yang, Jifeng Chen

1117

Intrusion Detection Based on Improved GA-RBF and PCA by Yuesheng Gu, Yanli Zhu, Peixin Qu

1122

Performance Analysis of Some Known Asymmetric Fingerprinting Schemes by Hongyan Wang, Yunyang Yan

1127

A Hybrid Price Forecasting Based on Linear Backfilling and Sliding Window Algorithm by Hong Zhou, Quanyin Zhu, Pei Zhou

1131

Implementation of LDPC Codes Decoding Based on Maximum Average Mutual Information Quantization by Lixin Li, Zhengkang Chen, Jie Fan, Nan Qi, Huisheng Zhang

1135

Perceptual Video Content Identification Based on Relative Orientation Invariant between Geometric Centroid by Dayong Wang, Yujie Zhou, Dandan Zhao

1140

Optimal Contention Window Adjustment for Asymmetry Traffic Over IEEE802.11 WLANs by Zhengyong Feng, Guangjun Wen, Yuesheng Gu

1145

(continued)


Sound Analysis for Diagnosis of Children Health Based on MFCCE and GMM by Chunying Fang, Haifeng Li, Wei Zhang, Bo Yu

1153

Edge Detection and Noise Reduction for Color Image Based on Multi-Scale by Feng Xiao, Mingquan Zhou, Guohua Geng

1157

Robust Collaborative Tracking in a Multi-Camera Surveillance System by Weichu Xiao, Weihong Chen, Xiongwei Fei

1163

Comparison between Three Algorithms for Smooth Support Vector Regression by Bin Ren, Huijie Liu, Lianglun Cheng

1169

PCNN-Histogram Based Multichannel Image Segmentation Algorithm by Beiji Zou, Haoyu Zhou, Geli Lv, Guojiang Xin

1175



Outline-Based Text Image Guorong Xiao1, Xuemiao Xu2

Abstract – Text image is a graphic design technique that consists of pictures pieced together from the printable characters. It is very useful in the currently text-based communication channels. Text image can be roughly divided into two major styles, tone-based and outline-based. Some programs allow one to automatically convert an image to text characters. However, they can only generate the tone-based text image as the tone-based one can be regarded as a simple dithering process. This paper propose a novel method to generate outline-based text image with the minimal number of characters, given the reference image and focus on the fixed-width characters used in traditional text image and ignore the proportional fonts. Convincing results are shown to demonstrate the effectiveness of the proposed method. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Text Image, Shape Similarity, Grid Deformation

I.

Introduction

Text image is a graphic design technique that uses computers for presentation and consists of pictures pieced together from the 95 printable characters defined by the ASCII Standard from 1963 and ASCII compliant character sets with proprietary extended characters [10]. The term is also loosely used to refer to text based art in general. Text image can be created with any text editor, and is often used with free-form languages. Text image is used wherever text can be more readily printed or transmitted than graphics, or in some cases, where the transmission of pictures is not possible. Text image is also used within the source code of computer programs for representation of company or product logos, and flow control or other diagrams [10]. For most benefit from broadband internet users may not realize the significance of the early development of text image. These simple design consisting of symbols, picture file for breaking the narrow bandwidth limitations, the visual creative writing through the same cable and quickly passed to the user's computer, it is not as graphic files, need special transcoding to read or even copy, whether it is a browser, telnet window, email application can read. It is a very free and open image form. Text image is no longer like you are familiar with the bbs as simple signature, not simply a few lines, some people will view it as a picture, so there is realism, cartoons, and even abstract creations. Some people treat it as a graphical authoring tool, an online literature into factions. Furthermore, it was text image into the static and dynamic film-like way, and later became the European avant-garde digital art. More recently, it was extended to the recent rapid development of mobile phone culture. Its development can be described as diverse and seems to have not less than hot. Manuscript received and revised October 2011, accepted November 2011

1084

Text image can be roughly divided into two major styles, tone-based and outline-based. Figs. 1 show one example.

Figs. 1. Text image: (a) Reference image. (b) Tone-based text image is generated by PicText, requiring a character grid of resolution 150*69. (c) Outline-based text image is manually created by an artist. The character grid resolution is 27*21 and significantly lower than (b)

The tone-based text image maintains the intensity distribution of the reference image (Fig. 1(b)), while the outline-based text image captures the 30 major structure of the reference image (Fig. 1(c)). It is obvious that outline-based text image uses much smaller number of characters and it is usually clearer than the tone-based one (Figs. 1(b) & (c)). To the best of our knowledge, there is no previous academic study on text image techniques. Instead, as a culture in the cyberspace, the best references of text image Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Guorong Xiao, Xuemiao Xu

can be found online. There is collaboratively prepared FAQ (frequently asked questions) for usenet newsgroup alt.ascii-art, which keeps track of the update information and resources related to text image. Other sources of reference are the online tutorials written by individual enthusiasts. Some research also attempts to automatically convert images to text 86 image. Some programs allow one to automatically convert an image to text characters, which is a special case of vector quantization. A method is to sample the image down to grayscale with less than 8-bit precision, and then assign a character for each value. However, they can only generate the tone-based text image as the tone-based one can be regarded as a simple dithering process. Note that the tone based text image normally consumes more text characters and the result may not fit into the limited text screen. In this paper, we focus on the generation on the outline-based text image as it presents a clearer picture in a smaller text space. Its generation can no longer be regarded as a dithering process. Instead, the shape and structure similarity should play a major role in its generation. In this paper, we propose a novel method to generate outline-based text image with the minimal number of characters, given the reference image. We focus on the fixed-width characters used in traditional text image and ignore the proportional fonts. The effectiveness of the method is demonstrated by examples.

II.

Our Framework

With the wide usage of text image, its automatic generation is desirable. However, existing methods can only handle the easier tone based text image, as its generation can be regarded as a dithering problem with characters. Satisfactory outline-based text images are mostly manually designed. As the text screens of modern mobile devices are limited, the character-saving outline-based text image is more practical for commercial usages. Unlike the tone-based text image reproducing the image tone, outline-based text image utilizes the shape of characters to approximate the image structure. However, due to the limited set of character shapes and restrictive placement of characters over the character grid, not all possible structure can be faithfully represented. Artists overcome the limitation by slightly deforming the reference image (or equivalently, the character grid) as demonstrated in (Fig. 1(d)).This explains the major challenge of the outline-based text image generation. We formulate the generation process as an optimization problem by minimizing the deformation of the character grid, the number of characters used, and the shape dissimilarity between the character shapes and the underlying image structure. The character grid deformation mimics how text image artists deform the image. Even after the deformation, the shape of characters and the underlying structure can be substantially different.


Hence, alignment sensitive metrics such as SSIM [9] is not applicable to measure the shape similarity. The translation, scale and rotational invariant shape context [1][6] is also not applicable as our application requires the selected characters to have a similar position, scale and orientation, as the underlying structure. We propose a novel shape similarity metric that can tolerate the misalignment while remain accounting for position, scale, orientation, and structure. The method of our framework, given an image (real photograph or cartoon), we start by obtaining its outline for the following generation of outline-based text image. This outline can be simply obtained by naive edge detection. Instead, we employ a more sophisticated line art generation method proposed by Kang et al. [4]. To focus only on the shape of the outline & character images and avoid the influence of line thickness, we perform thinning on the outline & character images so that all lines are with single pixel width [12]. The thinned image is further vectorized, so that it can be rasterized in arbitrary scale. As the limited shapes of text characters cannot represent all possible image content, artists slightly deform the image in order to allow the combination of characters to represent the deformed image. We mimic this by iteratively deforming a grid overlaid on the outline image. The initial grid is regularly laid. During each iteration, the current grid is deformed and the underlying image is rectified. Each grid cell content is mapped to a rectangular block and approximated by a best-matched character. An objective function is proposed to evaluate the current grid. An optimal grid is selected by minimizing the text resolution, the deformation of text grid, and the dissimilarity between the characters and the rectified image. The dissimilarity is measured according to a novel misalignment-tolerant shape similarity metric. Once the optimization is completed, we obtain the optimal text grid together with the associated text image. Due to the limited number of characters, it is very likely that the arbitrary image content cannot be well represented by the limited shapes of characters, even with the deformation. Therefore, we need a shape similarity metric that is tolerable to misalignment. Fig. 2(a) shows an imperfect circle positioned near the bottom. In our application, we can accept representing this imperfect circle with a character “o” even they are not well aligned (Fig. 2(b)). Moreover, we prefer the small letter “o” instead of the capital letter “O” (Fig. 2(c)) as its shape is closer to the underlying content. Existing shape similarity metrics can be roughly classified into two extreme categories, alignment -sensitive metrics and transformation-invariant metrics. PSNR (peak signal-to-noise ratio) or MSE (mean squared error), and the well-known SSIM [4] belong to the former category. Their similarity values drop significantly when two equal images are slightly misaligned during the comparison. On the other hand, the transformation invariant metrics are designed to be invariant to translation, scaling, or orientation.


1085


need to evaluate grids that are too large, because the beauty of outline-based text image is it fits into small text screen.

III. Results and Discussions Figs. 2. Matching image content with limited character shapes.(a) The image content of an imperfect circle. (b) Character “o” is a reasonable match in terms of shape and position, even it is slightly misaligned with the imperfect circle. (c) In contrast, capital letter “O” is less desirable, as its shape is much different from the image content

These metrics include shape context descriptor [1] [6], Fourier descriptor [11], skeleton-based shape matching [3] [7] [8], and curvature-based shape matching [2] [5]. In our application, we need, however, a metric that can tolerate misalignment while remain accounting for position, scale, orientation, and structure. To achieve this, we measure the shape similarity using shape features regularly distributed over the interested region. Given a shape positioned within an interested region, we regularly distribute N sample points. For each sample point, a shape descriptor is employed to quantify the local shape feature. We employ an efficient transformation-invariant metric, shape context [1][6], to evaluate the shape dissimilarity between the characters and the deformed image. Shape context is invariant to translation, scaling, rotation, and even allows small local perspective distortions. For each reference contour point, the shape context describes the distribution of the relative positions of all the other points in a spatial histogram. Based on shape context, we measure the shape distance between two shapes Sk and Sl by measuring the distances from Sk to Sl and from Sl to Sk , respectively [2]. This symmetric measurement is more stable on shape distance. The formulation is given by the following equation: D ( Sk ,St ) =

where M k

1 Mk

∑

i∈M k

fi − h j* +

1 Mt

∑

j∈M t

To validate our method, we test our method with some of input images. Our character database contains 314 distinct characters. They include 96 characters from ASCII table, 55 Greek characters, 90 Japanese Hiragana and Katakana, 64 Russian characters, and 9 table symbols (component lines to form table). Fig. 3(a) shows the “ox” after the one shown previously in Fig. 1(a). The text image generated by our method is shown in Fig. 3(b). Fig. 4 shows the “jingjing” generated by our method. It is the results of converting cartoon to outline based text image. Note that how the eyes in them are faithfully represented. The results of converting the calligraphy is shown in Fig. 5. Our current method does not handle proportional fonts. To extend our work to support proportional font, the grid layout has to be changed as the number of characters on a horizontal line is data-dependent and not fixed.

fi* − h j (1)

and M l are the total numbers of contour

( )

points on shape Sk and Sl , respectively; fi h j is the shape context feature vector for the i-th (j-th) contour point on shape Sk ( Sl ) . The j* −th contour point on Sl is the corresponding point to the i-th sample point on Sk , which is defined as j* = arg min j i* = arg min i

fi - h j . Similarly,

fi - h j . This measurement is simple in

computation and retains high accuracy. Since the shape feature vectors of all character images are precomputed, we avoid to resample the character images. We can simply raster the vector graphics in the desired resolution for matching. In practice, there is no


Figs. 3. Text image “ox”


1086


one can be regarded as a simple dithering process. We present a method that mimics how artists deform the given underlying image in order to approximate it by the limited variety of character shapes. Our method is formulated as an optimization to balance between the shape similarities, the text grid deformation. Some convincing results, that are comparable to manually prepared text image, have been shown to demonstrate the effectiveness of our method. We propose a novel shape similarity metric that can tolerate the misalignment while remain accounting for position, scale, orientation, and structure. Although the line art generation method is not our contribution, the quality of the generated line art does affect the resultant text image. Some convincing results, that are comparable to manually prepared text image, have been shown to demonstrate the effectiveness of our method.

Fig. 4. Text image “jingjing”

Acknowledgements This work was supported by National Natural Science Foundation of China(Grant No. 61103120), Guangzhou Novo Program of Science & Technology (Grant No. 0501-330), the Fundamental Research Funds for the Central Universities (Grant No. 2011ZZ0012), NSFC-Guangdong Joint Fund (Grant No. U1035004) ,and the National Natural Science Foundation of China (Grant No. 61070090 ).

References [1]

Fig. 5. Chinese calligraphy (“calligraphy”)

IV.

Conclusion

In this paper, we introduce the knowledge of text image, and do some research on the generation of text image. At present, some programs allow one to convert an image to text characters automatically, which is a special case of vector quantization. However, most of them can only generate the tone-based text image as the tone-based


Belongie, S., Malik, J., and Puzicha, J. Shape matching and object recognition using shape contexts. IEEE Transactions on Pattern Analysis and Machine Intelligience Vol. 24, n. 4, pp. 509– 522, 2002. [2] Cohen, I., Ayache, N., and Sulger, P. Tracking points on deformable objects using curvature information. In ECCV’92: Proceedings of the Second European Conference on Computer Vision, Springer-Verlag, London, UK(Page: 458–466, Year of Publication: 1992). [3] Goh, W.-B. Strategies for shape matching using skeletons. Comput. Vis. Image Underst. Vol. 110, n. 3, pp. 326–345, 2008. [4] Kang, H., Lee, S., and Chui, C. K. Coherent line drawing. In ACM Symposium on Non-Photorealistic Animation and Rendering (NPAR), pp. 43–50, 2007. [5] Milios, E. E. Shape matching using curvature processes.Comput. Vision Graph. Image Process.Vol. 47, n. 2, pp. 203–226, 1989.] [6] Mori, G., Belongie, S., and Malik, J. Efficient shape matching using shape contexts. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 27, n. 11, pp. 1832–1837, 2005. [7] Sundar, H., Silver, D., Gagvani, N., and Dickinson, S. Skeleton based shape matching and retrieval. SMI ’03:Proceedings of the Shape Modeling International (Page:130, Year of Publication: 2003). [8] Torsello, A., and Hancock, E. R. A skeletal measure of 2d shape similarity. Computer Vision and Image Understanding, Vol. 95, n. 1, pp. 1–29, 2004. [9] Wang, Z., Bovik, A. C., Sheikh, H. R., Member, S., SI Moncelli, E. P., and Member, S. Image quality assessment: From error visibility to structural similarity. IEEE Transactions on Image Processing, Vol. 13, pp. 600–612, 2004. [10] WIKIPEDIA, 2009. ASCII art. http://en.wikipedia.org/wiki/Ascii art.


1087


[11] ZAHN, C. T., AND ROSKIES, R. Z. 1972. Fourier descriptors for plane closed curves. IEEE Transactions on Computers 21, 3, 269–281. [12] ZHANG, T. Y., AND SUEN, C. Y. 1984. A fast parallel algorithm for thinning digital patterns. Commun. ACM 27, 3, 236–239. [13] Qingzhen Xu, Xianping Wu. The characterizations of optimal solution set in programming problem under inclusion constrains . Applied Mathematics and Computation, Vol. 198, Issue 1, pp. 296-304, 2008. [14] Kannala J,Rahtu E,and Heikkil J. Affine registration with multi-scale autoconvolution. International Conference on Image Processing,Genoa (Page:1064-1067. Year of Publication: 2005). [15] Rahtu E,Salo M,and Heikkil J. Multiscale autoconvolution histograms for affine invariant pattern recognition.Proc.16th British Machine Vision Conference,Edinburgh (Page: 1039-1048, Year of Publication: 2006) . [16] Petrou M and Kadyrov A. Affine invariant features from the trace transform. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 26, n. 1, pp. 30-44, 2004. [17] Kim T-K,Kittler J, and Cipolla R. Discriminative learning and recognition of image set classes using canonical correlations. IEEE Transactions on PAMI, Vol. 29, n. 6, pp. 1005-1018, 2007.


Department of Computer Science and Technology, Guangdong University of Finance. 2 School of Computer Science and Engineering, South China University of Technology.

Guorong Xiao received the B.E. and M.E. degrees in computer science and technology from South China University of Technology, Guangzhou City, China. He is currently a lecturer at the department of computer science and technology, GuangDong University of Finance, China. His research interests are animation, data warehouse and financial analysis etc.He has finished several banking, securities and mobile graphic applications. Xuemiao Xu is the Corresponding author, is now an associate professor in the South China University of Technology. Her research interests mainly include animation, digital Manga/Cartoon, real-time rendering and geometry compression.



1088


Integrated Price Forecast Based on Dichotomy Backfilling and Disturbance Factor Algorithm Quanyin Zhu1, Suqun Cao2, Pei Zhou1, Yunyang Yan1, Hong Zhou1

Abstract – In order to improve the accuracy of price forecast on imperfect data by web extracting, a novel repair data algorithm based on dichotomy backfilling is proposed in this paper. The price forecast algorithm based on the disturbance factor is utilized to verify the validity of repair dichotomy backfilling algorithm. Experiments demonstrated that the mean absolute errors only can be reduced 1.45 percent. Furthermore, the repair data algorithm based on average data backfilling and that based on dichotomy backfilling are explored as well to compare their accuracy. Experiments demonstrated the reduction of mean absolute errors in the price forecast verification model. The first one only can reduced 1.56 percent than the actual rate. But the second one can reduced 0.48 percent than the actual rate. Experiment results proved that this dichotomy backfilling algorithm is meaningful and useful to analyze and to research the price market on imperfect data by Web extracting. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Dichotomy Backfilling, Imperfect Data, Price Forecast Algorithm, Disturbance Factor

I.

Introduction

Forecasting is the process to make the statements about events whose actual outcomes (typically) that has not yet been observed. A normal sample data might be used to estimate for some variable of interest at some specified future date. Prediction is a similar way, but need more general term and correct data. Both of them might refer to formal statistical methods employing time series, cross-sectional or longitudinal data, or alternatively to less formal judgmental methods [1]. Predicts theory and method can be applied widely distributed in all kinds of areas of natural and social aspects. According to the areas covered, the different research objectives and tasks, forecasting can be classified with different areas of forecasting such as weather forecasting, scientific forecasting, military forecasting, technology forecasting, economic forecasting, and social prediction [2]. As we know, the Web mining can be used to discover the news [3], analyze email communications [4], and require skill sets for computing Jobs [5] and even to latent topics from web sites of terrorists or extremists [6] and so on. When shopkeepers want to know more information about the customer and recommend the new products to them using e-mails, but the most interested data is the commodities price. So, our group researched interesting on the commodities price evolution using Web mining for the shopkeeper’s selling online. Our team developed an application system to select the cell phone prices which are sold online.


1089

When we extract lots of data to build the cell phone prices data sets, more and more imperfect data are occur. Also we used those data to research the price forecast algorithm [7], [8] and the price dynamic trend analysis [9], but correct all the imperfect data become impossible issues when our data sets get more then one hundred thousand records. How to deal with those imperfect data and use it to forecast the price and reduce the error as much as possible is a valuable work. We use Price Forecast of Disturbance Factor (PFDF) to build the price forecast algorithm tested on delete some original data. After we verify the validity of Dichotomy Backfilling (DB) algorithm by comparison of Mean Absolute Errors Rate (MAER) on Original Data (OriD), MAER on Repaired the Data (RepD) and MAER forecasting on Mean Average Backfilling (MAB) algorithm respectively, we find the actual rate can be reduced very small.

II.

PFDF Algorithm

Some definitions used in this paper are given as follows. Single errors of predicted value: et = Yt − Yˆt , t = 1, 2 ," ,n

(1)

Relative errors of single predicted value: et =

et Yt − Yˆt = , t = 1, 2," ,n Yt Yt

(2)


Quanyin Zhu, Suqun Cao, Pei Zhou, Yunyang Yan, Hong Zhou

Mean Absolute Errors (MAE): MAE =

n

∑ et

1 n

=

t =1

1 n

n

∑ Yt − Yˆt

(3)

t =1

Here, y is the next average value of disturbance factors, and x is the value of weeks. In order to simplify the computing, let’s x=1, x=2, and x=3 into the equation respectively, we can get the ternary once equations:

Mean Absolute Percentage Errors (MAPE): 1 MAPE = n

n

∑ t =1

et

1 = Yt n

n

Yt − Yˆt

t =1

Yt

∑

⎧ y1 = a + b + c ⎪ ⎨ y2 = 4a + 2b + c ⎪ ⎩ y3 = 9a + 3b + c

(4)

(11)

That is: The t is the period predicted value, Yt is the actual value and Yˆ is the predicted value.

⎧a = ( y1 − 2 y2 + y3 ) / 2 ⎪ ⎨b = ( −3 y1 + 4 y2 − 3 y3 ) / 2 ⎪c = 3 y − 3 y + y 1 2 3 ⎩

t

ˆ . Set up tend line equation is Tt = aˆ + bt According to tend line equation, compute T1 ,T2 ," ,Tn in each phase. Reject tend: y St = t , t = 1, 2 ," ,n Tt

(5)

St =

Si + Si + L + Si + 2 L + " + Si + ( m −1) L m

Because of y1, y2, and y3 are known values. So we can get the value of a, b and c respectively, and express the equations (13) as: y1 − 2 y2 + y3 2 −3 y1 + 4 y2 − 3 y3 x + x+ 2 2 + ( 3 y1 − 3 y2 + y3 )

y=

Estimates disturbance factors then compute the average of the common season St , remove random disturbance. Set the average value as the disturbance factors estimates:

(12)

(13)

III. DB Algorithm Let time sequence data be A = { x1 ,x2 ," ,xn } , any

, i = 1, 2," ,L (6)

element is ∀xi ∈ A,i ∈ [1,n ] . Assuming that the time sequence data missed is B = {b1 ,b2 ," ,bm } ,any element is ∀b j ∈ B ⊂ A, j ∈ [1,m ] , and define missing data is

The sum of disturbance factors value should be equal to L, just: L

∑ i =1

Si = L

b j = xi . The missing data may be two situations.

(7)

But this kind of method gets the disturbance factors estimates value needs adjust. The adjust method is to find out in one phase, each disturbance factors estimates as regulation factor, like: S=

1 L

i =1

If m = 1 , then b1 = xi −1

or b1 = xi +1

(14)

If m = 2 , then

L

∑

The first is that B continues in A , the DB algorithm is as following:

Si

(8) b1 = xi −1 ,b2 = xi +1

Then using each phase disturbance factors estimates St divide regulation factor S, get the disturbance factors value: S (9) Si = i , i = 1, 2," ,L S

If m ≥ 3 , then If m mod 2 = 0 b1 ,b2 ," ,bm = x 2

2

y = ax 2 + bx + c

m i − +1 2

bm ," ,bm = xi + m

We can assume the equation of the average values is: (10)


(15)

(16)

+1

If m mod 2 ≠ 0 International Review on Computers and Software, Vol. 6, N. 6

1090


b1 ,b2 ," ,bm +1 2

−1

=x

i−

MAE on OriD is 1.85 percent, and the average MAE on RepD is 3.30 percent.

m +1 +1 2

bm +1 ," ,bm = xi + m 2

(17)

+1

TABLE I THE DETAIL OF ABANDONED DATA AND REPAIRED DATA Type Abandoned Data Repaired Data Time HTC A8180 3580,3580 3380,3380 9.19-9.20 Motorola 1999,1999, 2148,2148, 9.14-9.27 XT702 1999,1999 2148,2148 1548,1529,1529, 1548,1548,1548, LG P503 9.16-9.20 1529, 1529,1529 1529, 1529,1529 Nokia E66 1148,1148, 1148,1148, 7.6~7.10 1148,1148,1148 1148,1148, 1148

bm +1 = xi + m 2

The second is that B discrete in A , and then the missing data can be changed to the first situation, and using equation (14) to (17) procedure respectively.

IV.

Experiments

V.

We select the four type’s cell phone price extracted from www.360buy.com and www.dangdang.com at 2 July 2011 to 30 Sep. 2011. The equations (1) to (12) are used to verify the DB algorithm. The abandoned data and the RepD using equations (13) to (16) are listed in Table I, The results of PFDF on OriD compare with RepD are listed in Table II. The average

Algorithm Application

In order to deal with the imperfect data on the price forecast, we select four type cell phone price extracted from www.360buy.com and www.dangdang.com at 1 July 2011 to 30 Sep. 2011 to forecast the price. The detail is listed in Table III. The Average of MAE is 3.93 percent.

TABLE II THE PRICE FORECAST OF DISTURBANCE FACTOR ON ORIGINAL DATA COMPARE WITH REPAIRED DATA Type

Avg. (1)

Avg. (2)

Avg. (3)

Avg. (4)

Avg. (5)

Avg. (6)

Avg. (7)

Avg. (8)

Avg. (9)

Avg. (10)

Avg. (11)

Avg. (12)

Avg. Forecast Price of Oct.1 Price (13)

MAE

Original of HTC A8180

3783

3780

378

3754

3780

3783

3783

3780

3780

3780

3609

3494

3351

3323

3180

4.49%

Repaired of HTC A8180

3783

3780

378

3754

3780

3783

3783

3780

3780

3694

3580

3494

3351

3295

3180

3.59%

Original of MotorolaXT702

2127

2148

2148

2148

2148

2148

2148

2148

2148

2127

1999

1999

1999

2026

1999

1.33%

Repaired of MotorolaXT702

2177

2148

2148

2148

2148

2148

2148

2148

2148

2127

2064

2020

1999

2084

1999

2.35%

Original of LG P503

1699

1685

1649

1663

1639

1588

1549

1549

1576

1565

1539

1529

1525

1505

1529

1.57%

Repaired of LG P503

1699

1685

1649

1663

1639

1588

1549

1549

1576

1565

1534

1529

1525

1418

1529

7.26%

Original of Nokia E66

1182

1148

1142

1149

1147

1188

1188

1188

1188

1188

1188

1188

1188

1188

1188

0%

Repaired of Nokia E66

1217

1171

1142

1149

1147

1188

1188

1188

1188

1188

1188

1188

1188

1188

1188

0%

Avg.(i) is the weekly average price of i-th week data.

Type

TABLE III THE PRICE FORECAST OF DISTURBANCE FACTOR ON REPAIRED DATA USING DB ALGORITHM Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Avg. Forecast Price of Oct.1 Price (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)

Sony Ericsson X10i

3598

Motorola XT711 Sumsang I5088 Sumsang I809

MAE

3598

3484

3199

3199

3199

3199

3142

2799

2799

2513

2399

2399

2228

2339

4.76%

1999

1999

1999

1999

1985

1970

1899

1885

1799

1799

1799

1799

1799

1799

1799

0%

1082

1038

998

998

998

944

909

903

869

869

890

876

881

878

858

2.37%

4134

4022

4134

4134

3785

3770

3699

3628

3294

3719

3453

3176

3400

3449

3176

8.6%

Avg.(i) is the weekly average price of i-th week data.



1091


TABLE IV THE COMPARISON OF MEAN ABSOLUTE ERRORS RATE ON ORIGINAL DATA, ON REPAIRED DATA AND ON MEAN AVERAGE BACKFILLING Type

ActD

3380,3380,3380,3380, 3380,3380,3180 999,999,999,999,999, Motorola ME511 999,999 3639,3639,3750,3650, HTC T9191 3740,3750,3740 2999,2999,2999,2999, Motorola XT800 2999,2999,2999 1529,1529,1529,1529, LG P503 1529,1529,1529 MAER HTC A8180

VI.

ForD on OriD 3520,3500,3480,3460, 3440,3420,3400 1024,1024,1019,1014, 1009,1004,999 3642,3641,3640,3650, 3651,3661,3672, 2999,2999,2999,2999, 2999,2999,2999 1529,1529,1529,1526, 1526,1526,1526

MAER on OriD

1.50% 1.38% 0.00% 0.11% 1.25%

This work is supported by the National Sparking Plan Project of China (2011GA690190), the Major Program of the Natural Science Foundation of the Jiangsu Higher Education Institutions of China (11KJA460001); the Industry-University Collaboration Project of Huaian City, China (HAC201002); the fund of Huaian Industry Science and Technology. China (HAG2011052, HAG2011045).

References

[3]

[4]

[5] [6]

http://www.dictionary30.com/meaning/Forecasting Zhang Gui-xi and Ma Li-ping, An Introduction to Forecast and Decision. Beijing: Capital, Economic and Trade University Press, 2006. Cheng-Zhong Xu; Ibrahim, T.I., A keyword-based semantic prefetching approach in Internet news services, IEEE Transactions on Knowledge and Data Engineering, vol. 16 , n. 5, pp. 601 – 611, 2004. Grobelnik, M.; Mladenic, D.; Fortuna, B., Semantic Technology for Capturing Communication Inside an Organization, IEEE Internet Computing, vol. 13 , n. 4, pp. 59 – 67, 2009. Litecky, C.; Aken, A.; Ahmad, A.; Nelson, H.J., Mining for Computing Jobs, IEEE Software, vol. 27, n. 1, pp. 78 – 85, 2010. Hsinchun Chen, IEDs in the dark web: Lexicon expansion and genre classification, IEEE International Conference on


4.01% 3.03% 1.39% 0.00% 0.22%

MAB 3480,3460,3440,3420, 3400,3380,3380 1087,1082,1077,1072, 1067,1062,1057 3670,3669,3668,3678, 3665,3661,3672 3063,3063,3063,3063, 3050,3038,3025 1565,1565,1565,1554, 1547,1540,1533

1.73%

Conclusion

Acknowledgements

[1] [2]

3547,3527,3507,3487, 3467,3467,3413 1034,1039,1034,1029, 1024,1019,1014, 3645,3643,3642,3652, 3652,3661,3672 2999,2999,2999,2999, 2999,2999,2999 1535,1533,1531,1526, 1526,1526,1526

3.27%

The PFDF is used to verify the validity of DB algorithm. The MAPE only can be reduced 1.45 percent. We used this algorithm on the application system and achieved excellent effects. Furthermore, we used the five type’s cell phone price extracted from same Web site and the same time to forecast the price. The comparison results of MAER on OriD, MAER on RepD and MAER on MAB are listed in Table IV. We can find that the MAER of forecasting price on MAB method increases 1.56 percent than the actual rate. But the MAER of forecasting price on DB algorithm only increases 0.48 percent than the actual rate. The price extracting and forecasting are obtained more attention by the shopkeepers who can use them to sell online. Our future interesting work is to build the data sets, to look for more effectual algorithms, to forecast price.

MAER on RepD

RepD

[7]

[8]

[9]

MAER on MAB 2.20% 7.30% 1.38% 1.38% 1.83% 2.81%

Intelligence and Security Informatics (Page: 173 – 175 Year of Publication: 2009 ISBN: 978-1-4244-4173-0). Quanyin Zhu, Suqun Cao, Jin Ding and Zhengyin Han, Research on the Price Forecast without Complete Data based on Web Mining, 10th International Symposium on Distributed Computing and Applications to Business, Engineering and Science (Page: 120 – 123 Year of Publication: 2011 ISBN: 978-0-7695-4415-1). Quanyin Zhu, Yunyang Yan, Jin Ding and Jin Qian, The Case Study for Price Extracting of Mobile Phone Sell Online, The 2nd IEEE International Conference on Software Engineering and Service Sciences (Page: 282 - 285 Year of Publication: 2011 ISBN:978-1-4244-9698-3 ). Quanyin Zhu, Hong Zhou, Yunyang Yan, Jin Qian and Pei Zhou, Commodities Price Dynamic Trend Analysis Based on Web Mining. Third International Conference on Multimedia Information Networking and Security(Page: 524 - 527 Year of Publication: 2011 ISBN:978-0-7695-4559-2 ).


Faculty of Computer Engineering, Huaiyin Institute of Technology, Huaian 223003, Jiangsu, China. 2 Faculty of Mechanical Engineering, Huaiyin Institute of Technology, Huaian 223003, Jiangsu, China.

Quanyin Zhu received the B.S. degree from University of Electronic Science and Technology of China, Chengdu, Sichuan, China, in 1990. He was a Visiting Scholar in the Department of Automatic Control of Southeast University, Nanjing, Jiangsu, China, in 2000-2001, and International Visiting Scholar in school of Computing, Engineering and Information Science of Northumbria University, Newcastle, UK, in 2007. Mr. Zhu is currently a Professor at the Faculty of Computer Engineering, Huaiyin Institute of Technology, Huaian, Jiangsu, P.R.China. His current research interests include data mining, intelligent information processing, and software engineering. He is the author or coauthor of more than 20 papers published in international journals and conferences. He is the senior reviewer of the Journal of Huaiyin Institute of Technology. Suqun Cao was born at Huai’an City, Jiangsu Prov., P.R.China, in 1976. He received the B.S. degree in Mechanical Engineering from Jilin University of Technology, P.R.China, in 1997, the second B.S. degree in Software Engineering from Tsinghua University, P.R.China, in 2002, the M.S. degree in Computer Science from Southeast University, P.R.China in 2008 and the Ph.D. degree in Computer Science from Jiangnan University, P.R.China,


1092


in 2009. Since 2010 he has been an associate professor at the Faculty of Mechanical Engineering, Huaiyin Institute of Technology, P.R.China. His current research interests include machine learning, pattern recognition and intelligent fault diagnosis. Pei Zhou was born Yancheng City, Jiangsu Prov, P.R.China, in 1990. He is currently an Undergraduate student at the Faculty of Computer Engineering, Huaiyin Institute of Technology, Huaian, Jiangsu, P.R.China.

Hong Zhou received the degree of Master of Science from Abertay Dundee University, Dundee, Scotland, UK, in 2008. She is currently a Lecturer at the Faculty of Computer Engineering, Huaiyin Institute of Technology, Huaian, Jiangsu, P.R.China. Her current research interests include Web extracting and mining, Data warehouse and text data mining.

Yan Yunyang (1967-), received the B.S. degree from Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu, China, in 1988, received the M.S. degree from Southeast University, Nanjing, Jiangsu, China, in 2002, received the Ph.D. degree from Nanjing University of Science and Technology, Nanjing, Jiangsu, China, in 2008. He is currently a Professor at the Faculty of Computer Engineering, Huaiyin Institute of Technology, Huaian, Jiangsu, P.R.China. His current research interests include image processing, pattern recognition and applications of soft computing. He is the author or coauthor of more than 30 papers published in international journals and conferences. He is the senior reviewer of the Journal of Huaiyin Institute of Technology and Journal of Computational Information Systems.



1093


A Target Location Method Based on Binocular Stereo Vision for Eggplant Picking Robot Jian Song

Abstract – A target location method based on binocular stereo vision is proposed in order to provide the eggplant picking robot with the three-dimensional information of the fruit target. The camera imaging geometrical model is built for the self demarcation of the binocular vision system. The brightness-based threshold segmentation algorithm is adopted to segment G-B grayscale images. Such features as the contour profile, area, centroid, enclosing rectangle, cut-off point of the fruit target are extracted. The centroid is selected as the matching primitive and The algorithm with epipolar constraint, uniqueness constraint and disparity gradient constraint is adopted for the implementation of the fruit target matching. The eggplant fruit depth information is calculated in accordance with the homothetic triangle theory. It is measured through the experiment that the errors of the eggplant depth information mainly range within ±20 mm in the measurement distance of 300mm~600mm with the average time used 0.36s. The target location method based on binocular vision of the eggplant picking robot has simple principle and wide adaptability, and can be able to meet the requirements for target location of the picking robot. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Image Processing, Binocular Vision, Target Location, Picking Robot

I.

Introduction

In 2009, Chinese vegetable planting covered an area of 273 million mu(approximately 19 million hectares) which is 43% of the total sown area in the world, with a total output of 602 million tons which is 49% of the total output in the world, whose gross product outnumbered 880 billion yuan. Harvesting or picking is the most effort-requiring and time-consuming link in vegetable production operation, which, according to statistics, approximately accounts for from 40% to 60% of all the working out [1],[2]. Moreover, it requires picking in good time to guarantee the product quality. Meanwhile, the quality of the picking operation has an immediate influence on the quality, storage and follow-up process of the harvested eggplants, thereby finally affects the market price of eggplants and the ultimate benefit of the vegetable growers [2] [3]. Owing to the complexity of picking operation, the degree of picking automation is still very low. At present, the domestic fruit and vegetable picking operation is basically done manually. With the aging of population and the decrease of farming labor force, agricultural production cost is increased correspondingly, which greatly lowers the competitive power of the products. Therefore, it is of great significance to develop mechanized harvesting technology and to research on the vegetable picking robots. At present, color camera is the frequently-used target detection tool for agricultural robot. However, one color camera can only acquire the two-dimensional positioning


1094

information and the maturity information of the target rather than the three- dimensional position information. In addition, one single color camera is difficult to guarantee the accuracy and reliability of the information. Now the picking robots at home and abroad usually adopt binocular machine vision or monocular vision plus proximity sensor to achieve fruit positioning. Yet the precision of the proximity sensor is not high in the required fruit distance range, therefore, the research on binocular vision is more extensive. The vision system of the apple-picking robot developed by Teruo Takahashi of Hirosaki University in Japan in 2002 mainly adopted binocular stereo vision system composed of two color cameras. When the left and right camera simultaneously acquire the image of the same target, central composite of the two images are conducted to reestablish the three-dimensional information of the picking target [7]. Jiang Huanyu studied the identification and positioning method of ripe tomatoes based on binocular stereo vision technology and obtained positional information of the ripe tomatoes for the guidance of the automatic picking operation in the greenhouse [8]. Yuan Ting etc., achieved stereo matching and three-dimensional reconstruction of the grab point for cucumber in binocular vision through establishing matching strategy based on the combination of gray correlation and epipolar geometry [9]. Lv Xiaolian, et al., used information of color feature differences between the ripe tomato and the background for image segmentation to identify the ripe tomato; on the basis of camera calibration Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Jian Song

and centroid matching, space position information of the tomato fruit is acquired through 3-D reconstruction. Zhou Jun adopted Hough transformation in the image space to detect the fruit target and used 3-D coordinate information of multiple points of uniform distribution on the random sampling target to reconstruct fruit sphere model and then obtained space coordinates of the target centroid. Binocular vision technology has developed fairly fast in the application of picking robot, but it still has the problems of low accuracy of identification and positioning and work efficiency. In this paper, with the eggplant in growing environment as the research object, proper algorithm is used to segment the ripe eggplant image in accordance with the color features of the ripe eggplant to identify the target object from the background, and the positional information of the eggplant target is calculated with the algorithm based on centroid matching and supplemented by epipolar constraint, uniqueness constraints and disparity gradient constraints. This method can identify and position quickly the ripe eggplant so as to improve the intelligence and real-time of the vision system.

II.

1) Use AutoCAD cartographic software to draw a checkerboard grid diagram and print it with a laser printer, and then stick it on an ideal hard plate. 2) Move the camera by hand. Each of the two cameras collects 12~30 images of checkerboard grid diagram from different angles at the same time. 3) Detect all the grid intersection points in the image acquired. 4) Establish the camera imaging geometrical model, and use the maximum likelihood to estimate the internal parameters and external parameters of the camera. The result is shown in Fig. 1. 5) Use the external parameters obtained in the above step to solve the relative positioning problem or external positioning problem of the two cameras.

Calibration of Vision System

The basic task of the computer vision is to start from the image information acquired by the camera to calculate the geometric information of the object in 3-D space and reconstruct and identify the object. The correlation of the 3-D geometric position of certain point on the surface of the object and its corresponding point in the image is determined by the geometrical model of the camera image formation. The geometrical model parameters (the camera parameters) include the camera internal parameters and external parameters. In most cases, these parameters must be obtained through experiments and calculation. This course is called camera calibration. On the whole, the existing camera calibration techniques can be boiled down to two categories: the traditional camera calibration technique and camera self calibration technique. The traditional camera calibration is to calibrate by using the calibration object whose geometric dimensions in 3-D space are accurate and known. It can generally achieve good result but requires expensive equipment and complicated setting. Camera self calibration technique does not need any calibration object but only moving the camera in the stationary scene, which offers two constraint of the camera internal parameters. Therefore, the camera internal parameters and external parameters can be obtained from three images of the same scene shot by the camera with fixed internal parameters, which can be used very simply for 3-D information reconfiguration. The camera self calibration technique is adopted in this paper for vision system calibration. The concrete steps are as follows:


Fig. 1. The result of Calibration of vision system

III. Identification of Eggplant Fruit The task of the vision system of the picking robot for fruit and vegetable is to identify the ripening fruit from the complex background in accordance with the colors, sizes and shapes of the target fruits. So the characteristic that different position of the crop has different color can be used to identify the plant and its fruits. By the statistic analysis of the color features of different objects (fruit, branches and leaves, other background) in the growing environment through experiment, simple and practicable identification model is discovered to segment the image and extract the features of the target fruit. III.1. Image Segmentation Image segmentation in picking robot vision system is to extract and mark the eggplant fruits from the image, i.e., to segment the image into two parts: the eggplant fruits and the background. In line with the color feature analysis of the eggplant fruits and their surroundings, G-B grayscale images are most favorable for segmentation of the fruit target. In addition, the requirements for the real time and adaptability of the vision system of the picking robot are considerably high. In view of the two points above, brightness-based threshold segmentation algorithm is adopted to segment the G-B grayscale images.


1095

Jian Song

Due to the complexity of the eggplant growing environment, it is quite difficult to acquire preferable segmentation effect by the fixed-threshold-based segmentation method, while it is hard for such automatic threshold segmentation methods as OSTU, iteration method, Minimum error probability method to adapt to the picking robot’s requirements with a relatively slow segmentation speed and an unsatisfactory timeliness. On account of the above considerations, the threshold segmentation algorithm based on brightness is put forward which can realize the ideal segmentation of the eggplant fruit target. The main principle is: Th = Gav + ( Gmax − Gav ) f

(1)

where: Th is the segmentation threshold value; Gav is the mean gray value; Gmax is the maximum gray value; f is the weight factor. Through segmentation experiments on 30 eggplant images, f values range from -0.5 to 0.5. In general, when f = 0.1, it can get preferable segmentation effect. The image segmentation effect is shown in Fig. 2.

Point cut-off

Grab point center of gravity

Enclosing rectangle

Fig. 3. The result of feature extracting

The distance between the two cameras in the x direction is baseline distance B. In the model, the same feature point in the scene has different imaging positions in the two camera image planes. The difference of the two positions becomes parallax ( xl′ − xr′ ). The projective points of the feature point on the two different images are called conjugate pairs. The plane via the two camera centers ( Ol and Or ) and the scene feature point P is called epipolar plane. The intersecting line of the epipolar plane and the image plane is called epipolar line. All the epipolar lines on the same image plane intersect at one point which is called the epipolar point.

Fig. 2. The result of eggplant image segmentation

III.2. Extraction of the Fruit Target Features Image feature refers to the characteristics possessed by certain object in the image. For vegetable and fruit picking operation, it ultimately demands robot visual system to provide feature parameters of the picking target, such as the centre of gravity, the area, Minimum Enclosing Rectangle (MER), point cut-off, etc. Through image segmentation, binary image of the fruit target is acquired, for which edge extraction and contour tracing are dealt with, and the contour is labeled in two-dimensional array. In this way, it is convenient to extract the features desired for the target fruit. Fig. 3 is the rendering of feature extraction.

IV.

Binocular Vision Positioning Algorithm IV.1. Binocular Vision Ranging Principle

The fundamental solid geometrical relationship of binocular vision system is shown as in Fig. 4. It is composed of two identical cameras with the two image planes located in the same plane. The coordinate axes of the two cameras are parallel with each other with the coinciding x-axis.


Fig. 4. The geometrical model of Binocular vision distance measurement

In Fig. 4, the projective points of the scene point P on the left and right image plane is respectively Pl and Pr , which is without loss of generality, provided that coordinate system origin coincides with the left lens center. It is be prone to get depth information z according to the homothetic triangle theory: Z=

BF x'l

− x'r

(2)

where: F is the focal distance and B is baseline distance B. IV.2. Centroid Matching Algorithm Stereo matching is to establish the corresponding relationships of the features to correspond the pixels of the


1096

Jian Song

same spatial physics point in different images on the basis of the calculation of the selected features. The picking object is eggplant fruit which is cylindrical. The centroids of the images acquired by the two installed cameras in parallel are a good pair of matching points. Centroid matching method can be applied for positioning the target object. The coordinates of the centroid of the stereo-image pair of the target object can be obtained through image segmentation of the left and right images. And then the algorithm of epipolar constraint, uniqueness constraint and disparity gradient constraint is adopted to achieve the fruit target matching. Epipolar constraint gives the important constraint conditions of the matching point, which can narrow down the search space of the matching point, thereby improve the search speed and reduce the mismatches. From analysis and derivation, the relative position and extreme formula of the two cameras is: −1 u2T [ m ]x M 21M 11 u1 = 0

The feature extraction method introduced previously is used to extract the centroid coordinates of the eggplant fruit target, and then the algorithm with epipolar constraint, uniqueness constraint and disparity gradient constraint is employed to implement the fruit target matching. The specific algorithm is as follows: 1) Extract the centroid coordinates of the fruit target in the left and right images. 2) For each feature of the left image, use epipolar constraint conditions to search for the possible matching in the right image. 3) Use uniqueness constraint to match centroid coordinates of the eggplant fruit in the left and right images according to the matching theory; 4) Calculate the disparity gradients of each such pair to ensure that they do not violate disparity gradient limit. 5) End when all the possible matches are extracted.

(3)

V.

where: [ m ]x is anti-symmetric matrix defined by the vector m ; u1 and u2 are the homogeneous coordinates of the corresponding points of the two cameras; M 21 and M 11 are the camera projection matrixes. This constraint refers that in general, each feature point on one image can only correspond with at most one feature point on the other image. In the study, the eggplant centroid coordinates extracted separately in the two images are match in line with that one eggplant has only one centroid to meet uniqueness constraints. Disparity gradient measures the relative parallax. Suppose two points A and B in the space is

( Ar = ( axr

IV.3. Algorithm Implementation

) and Bl = ( bxl by ) in the left image while a y ) and Br = ( bxr by ) in the right image, the

Al = axl a y

Experiment Results and Analysis V.1.

Experiment Conditions

The eggplant picking robot with an open architecture is composed of PC computer, motion controller, AC servo, machine vision system and the robot body, as shown in Fig. 1. The machine vision system includes PC computer with Intel 2.4G CPU, two Zino DH-CG320 and Panasonic WV-CP470 color cameras, which are fixed on the manipulator base with a distance between of 600mm.The optical axes of the two cameras are parallel and of the same height. They are controlled by a trigger to shoot simultaneously images of the same scene with the image resolution ratio 640×480 pixels. In this paper, the image preprocessing, image segmentation, stereo matching and the calculation of the target object depth value are achieved through programming under Visual C++6.0 environment.

one-eyed separation degree S of the two points is:

(

1 S ( A,B ) = ( xl + xr )2 + ax − by 4

)

V.2. 2

(4)

The matching parallax difference is: D ( A,B ) = ( axl − axr ) − ( bxl − bxr ) = xl − xr

(5)

In this paper is the parallel binocular stereo vision system, so ay=by, the disparity gradient of the matching pair is: D ( A,B ) 2 ( xl − xr ) (6) Γ ( A,B ) = = S ( A,B ) ( xl + xr ) In practical application, disparity gradient is restricted: Γ ( A,B ) ≤ 1

Result and Analysis

Centroid matching algorithm is used for positioning experiment of 3 different eggplants on 10 different positions in the range of 300mm~800mm. The actual values of the depth information of the 3 eggplants on different positions are measured, and the distance values of the target objects are calculated with the centroid matching method, with the test result shown in Fig. 5, in which the ordinate is the deviation between the measured value and the predicted value of the eggplant depth information, while abscissa is the measured distance. It can be observed from the chart that when operating distance is less than 600 mm, the errors of the eggplant depth information mainly range between ±20 mm except for several singular points, and the time is approximately 0.07 s needed for positioning measurement of the target object for one time.

(7)



1097

Jian Song

measurement error

Acknowledgements This work is supported by Shandong Provincial Natural Science Foundation, china (No.Y2008G32) and Shandong Provincial universities Scientific Research Project (No.J09LG53).

References

Actual Distance(mm)

[1]

Fig. 5. Curve of measure distance

The main causes of the errors are: 1) Although the segmented image is filtered, it cannot be radically eliminated, which causes the errors of the centroid coordinates and consequently affects the accuracy of the depth information. 2) The reason that the measurement error increases with the increase of the distance is that disparity of the target object in the image decreases with increase of the distance; in cases of the same disparity errors, the errors of the predicted values increases with the increase of the distance. 3) Errors in camera demarcation also affect the measurement accuracy to some extent. It is demonstrated through the experimental results that the target location method based on binocular vision of the eggplant picking robot has simple principle, good intelligence and wide adaptability, and can be able to meet the requirements for target positioning of the picking robot except that it is time-consuming and the processing speed needs to be further improved.

VI.

Conclusion

A target location method based on binocular stereo vision is proposed in order to provide the eggplant picking robot with the three-dimensional information of the fruit target. As there are big differences between the ripe eggplant fruit and the background in RGB color space, brightness-based threshold segmentation algorithm can be adopted to preferably segment the target fruit. The algorithm based on centroid and supplemented with epipolar constraint, uniqueness constraint and disparity gradient constraint can be used to obtain accurately the three-dimensional information of the eggplant fruit. It is measured through the experiment that the errors of the eggplant depth information mainly range between ±20 mm in the measurement distance of 300mm~600mm with the average time used 0.36s. Though the method can be able to meet the requirements of the picking robot vision system, there is further room to improve the measurement accuracy and real-time. In future research, it needs to optimize the system algorithm design and demarcate accurately the camera parameters, thereby to develop the fruit target location algorithm that can be applied to the practical operations of the picking robot.


J. Song, T. Z .Zhang, L. M. Xu, Research Actuality and Prospect of Picking Robot for Fruits and Vegetables, Transactions of the Chinese Society for Agricultural Machinery, vol. 37, n. 5, pp. 158-162, 2006. [2] X. Y. Tang, T. Z. Zhang, Robotics for Fruit and Vegetable Harvesting: a Review, Robot, vol. 27, n. 1, pp. 90-96, 2005. [3] J. Song, T. Z. Zhang, B. Zhang, Development and experiment of Opening Picking Robot for Eggplant, Transactions of the Chinese Society for Agricultural Machinery, vol. 40, n. 1, pp.142-145, 2009. [4] K. l. Zhang, Y. Li T. Z. Zhang, Object Locating Method of Laser-assisted Machine Vision for Strawberry-harvesting, Transactions of the Chinese Society for Agricultural Machinery, vol. 41, n. 4, pp. 151-156, 2010. [5] Liu Zhaoxiang, Liu gang, Qiao Jun, Laser ranging sensor for apple harvesting robot”, Journal of Jiangsu University, vol. 31, n. 4, pp. 373-377, 2010. [6] Lu Xiaolian, Lu Xiaorong, Lu Bingfu, Design and research on tomato-harvesting robot visual system, Transducer and Microsystem Technologies, vol. 29, n. 6, pp.21-24, 2010. [7] P. R. Apeagyei, Application of 3D body scanning technology to human measurement for clothing Fit, International Journal of Digital Content Technology and its Applications, vol. 4, n. 7, pp. 58-69, 2010. [8] S. D. Ghode, S. A. Chhabria, R. V. Dharaskar, Computer Interaction System using Hand Posture Identification, International Journal of Digital Content Technology and its Applications, vol. 4, n. 5, pp. 21-33, 2010. [9] Zhou Jun, Liu Rui, ZhangGaoyang, Design ofFruitHarvesting RobotBased on Stereo Vision, Transactions of the Chinese Society for Agricultural Machinery, vol. 41, n. 6, pp. 158-162, 2010. [10] R. Niese, A. Al-Hamadi, A. Panning, D. Brammen, Towards Pain Recognition in Post-Operative Phases Using 3D-based Features From Video and Support Vector Machines, International Journal of Digital Content Technology and its Applications, vol. 3, n. 4, pp. 21-33, 2009. [11] Shigehiko Hayashi, Katsunobu Ganno, Yukitsugu Ishii., “Robotic Harvesting System for Eggplants”, JARQ 2002, vol. 36, n. 6, pp. 163-168, 2002. [12] N. Kondo, M. Monta, T. Fujiura, Fruit harvesting robot in Japan, Adv. Space Res, vol. 18, n. 2, pp. 181-184, 1996.

Authors’ information College of Machinery, Weifang University, Weifang 261061, Shandong, China. Jian Song received the Ph.D. degree in agricultural mechanization engineering from China Agricultural University, China in 2006. Currently, he is a associate professor at Weifang University, China. His research interests include image processing and Agricultural robot technology.


1098


Credit Card Risk Detection Based on Chaos Theory and Cloud Model Yanli Zhu, Yuesheng Gu, Shiyong Li, Shunping Wang

Abstract – Detecting fraud transactions has been a commonly concerned problem in the credit-card industry at present, and it directly affects the development of credit card business. To solve the problems of existing prediction methods of high false-positive rate, a new detection method of fraudulent transaction based on chaotic time series has been proposed. The method, firstly, tries to implement a dynamic customer segmentation method based on cloud theory to classify the customers properly. Then it tries to understand the nature of short-term credit card transaction through time series chaotic analysis and constructs predict model by employing back propagation neural network, to detect fraudulent transactions for each kind of clients. The experimental results with real data demonstrate that the method has lower space complexity and higher prediction accuracy than the BP neural network prediction. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Risk Detection, Chaos Theory, Phase-Space Reconstruction, Cloud Model, Dynamic Customer Segmentation

I.

Introduction

Higher asset earning rate of credit card has brought large profit for offers. It is also considered as a kind of high-risk business because credit card fraud costs cardholders and issuers hundreds of millions of dollars each year. How to detect credit card fraud transactions effectively, rapidly and exactly has been a commonly concerned problem in the financial industry at present [1] [2]. If detection system can be set up to pick out risk-taking, warning signals will be sent out to the monitors. Therefore, appropriate measures can be taken to minimize the loss. Many experts and scholars, using different methods, have proposed their model in risk prediction of credit-card transactions, such as statistical model, Bayesian and neural networks, etc [3] [4]. In [5], the author has presented that fraud warning system can be established by using genetic algorithms. Customer's consuming behavior possesses the main properties: the uncertain, nonlinear and chaotic property, which is always affected by many factors such as the mood. It is hard to establish an accurate mathematical model to describe. Therefore, predictive accuracy of the above model is very low, and it can’t fit the actual requirement. As a deterministic nonlinear dynamics system, chaotic system is extremely sensitive to the initial conditions. Chaos limits predictability so that the long-term prediction of time series is very difficult. Back propagation neural network has been the most widely used prediction model in many subject areas. Moreover, many researchers in various fields for the


1099

regularity of the system have achieved the good effect assuming the artificial neural network advantage in dealing with the nonlinear mapping [6]-[9].The paper [10] proposed a new outlier detection method based on chaotic behavior prediction and constructed an approximator for financial operations by employing radial basis function neural network. Experiments on synthetic data mixed with real-world data, and simulated outlier data demonstrated encouraging performance in outlier extraction. In this paper, we adopt the BP artificial neural network and chaotic theory to judge whether the current trade is fraud. It based on the fact that although cheating can steal the secret information of credit card, it is hard for them to imitate the history behavior of the real cardholder, which can be concluded from much trade information about trading time, place, amount, business category, frequency and so on. First of all, credit card customers should be classified according to their consumption characteristics. Then we calculate the optimal embedding dimension and time delay parameters with the C-C method before being processed by the phase space reconstruction for each kind of clients. Finally, we construct to predict model by employing back propagation neural network. The experimental results indicate that it is an effective way to improve the prediction accuracy and decrease the rate of false alarms by using chaos theory and BP.

II.

Cloud Model

Cloud model, put forward by Prof. Deyi Li, is a model of mathematical method, which is used to process the


Yanli Zhu, Yuesheng Gu, Shiyong Li, Shunping Wang

uncertainty and randomness. It is based on fuzzy logic and probability, providing a means of between qualitative and quantitative [11]. The cloud is composed of many of the cloud drops. A drop is a one-time value of qualitative concept. The total cloud reflects the radical character of fuzzy concept. U (the universe of discourse) is defined as U= {u}. T is the language value related to U. The degree of pertain which is called as CT ( u ) , is the degree that u belongs to U. The number feature of cloud is signified with the expected value, entropy and hyper entropy. So the cloud model is recorded as A ( Ex,En,He ) . The expected value Ex is the center value of the concept. The entropy En is defined by the bandwidth of the MEC (mathematical expected curve) of the normal cloud showing how many elements in the universe of discourse could be accepted. The MEC of the normal cloud is defined by the following formula: MEC A ( u ) = exp −

( u − Ex )2 2 En 2

(1)

series and reconstruct the phase space x ( k ) of the time

The entropy He is the uncertain degree of cloud drop. If the number characteristics Ex, En and He of the cloud are given, the following algorithm can be used to produce normal cloud: 1) The normal random number xi is calculated by the following formula: xi = G ( Ex ,En )

and understanding of the world, chaos theory has gradually become an important theory in complex systems, which has been widely applied in many fields, such as predicting seizure of falling sickness, prediction of financial markets, production system modeling, weather forecast and fragment geometric figure structure. Packard and Takens have proposed phase-space reconstitution theory[12], which introduces chaos theory into the time-series analysis. Chaos forecasting is established based on phase-space reconstitution of time series. Reconstituting proper phase-space and making the space orbit fully spread is the premise of further chaos series forecasting. The basic idea of the state space reconstruction is that a signal contains information about underserved state variables, which used to predict the present state. The process of reconstruction is determined by a choice of the optimal parameters τ and m. We calculate the delay time τ and the phase space minimum embedding dimension m with the C-C method [13]. Suppose { x ( k ) ,k = 1, 2 ,… ,N } is the detection time

(2)

2) Then the normal random number En' i is created with

series according to Takens’embedding theorem:

( ) X 2 = ( x ( 2 ) ,x ( 2 + τ ) ,...,x ( 2 + ( m − 1)τ ) ) X 1 = x (1) ,x (1 + τ ) ,...,x (1 + ( m − 1)τ )

(5)

...... X N = ( x ( N ) ,x ( N + τ ) ,...,x ( N + m − 1)τ )

where τ is the delay time; m is the embedding dimension; N = n − ( m − 1)τ is the effective length of vector series.

the entropy En and the standard deviation He by the following formula:

IV. En' i

= G ( En ,He )

(3)

3) (xi，µi) is viewed as cloud drop, µi is calculated by the following formula:

µi = exp −

( xi − Ex )2 2 E' 2ni

(4)

The above algorithm can be implemented in both hardware and software, which is named as cloud generator. The combination of the two kinds of generators can be used interchangeably to bridge the gap between quantitative and qualitative knowledge.

Chaos-BP Prediction Model

Chaos time sequence has an internal certain regularity which is nonlinear. It shows the correlation of time sequences in state space of time delay. This feature enables the system performs some kind of memory, but it is difficult to express the regularity by usual analysis method. Neural Network just has the right way of this kind of information processing. According to Kolmogorov continuity theorem and Takens theorem, the main process is used to set up a chaos-BP prediction model as follows [14]: Step 1. Reconstruct an m-dimensional phase space for a given chaos time sequence x(t) by selecting appropriate delay parameter and embedding dimension m: Y ( t ) = x ( t ) ,...,x ( t + ( m − 1)τ ) ∈ R m , ( t = 1, 2 ,.....) (6)

III. Phase-Space Reconstitution Chaos theory is the study of complex, nonlinear, dynamic systems. With the development of the science


Step 2. Regard m sample points in the phase space as m inputs of neural network, and establish a Chaos-BP prediction model. Take: International Review on Computers and Software, Vol. 6, N. 6

1100


xt [i ] = x ( t + ( i − 1)τ ) , i = 1, 2 ,...,m; t = 1, 2 ,...

(7)

as the network input, and we can get: ⎛ m ⎞ Z t = ft ⎜⎜ ∑ ( wt [i ] xt [i ]) ⎟⎟ ⎝ i =1 ⎠

(8)

Note: Z t is the output node of neural network, wt [i ] is the connection weight from i-th node of input layer to Z t node of output layer, ft is a nonlinear sigmoid function: 1 ft ( x ) = (9) 1 + e− x After establishing the Chaos-BP prediction model, we can simulate and predict the given chaos time sequence by using the neural network method. In practice we input: x ( t + 1) ,x ( t + 1 + τ ) ,...,x ( t + 1 + ( m − 1)τ )

into neural network, and the model will output x ( t + 2 ) + ( m − 1)τ .

So we can get a prediction result of given chaos time sequence by employing BP neural network.

V.

Predication Method Based on Chaos Theory and BP Neural Network V.1.

Data Preprocessing

Because the credit card system database is application oriented rather than subject oriented, the first step is to select the data relevant to the subject [15]. There are 100 tables of the data sheet in the credit card database, including customer application data, account information, business data, credit card transaction data flow, stop-payment list data, high-quality customer data. After analysis, we selected a few tables, which are related to the theme from these tables as follows: the personal customer information table, the card information table, the transaction logged table, overdraft history table, the balance of history table. For reflecting the user’s behavior model in the past trade, some statistical fields indicating the transaction information are added in the paper. The temporary table is formed according to the calculation on card trading log table and overdraft history table. In our experiment, we transformed the credit card data to format of data mining, mainly by the following means, such as, data cleaning, data conversion, data integration, and data discretization. There are 2180 credit customers, whose transaction data we choose from January 5 to December 30 in 2003, including 141562 trading records and their relevant information. For various reasons, the original database


has a lot of dirty data, such as incorrect attribute values, duplicate records, null value, inconsistent values, various abbreviations, violations of referential integrity and so on. In order to make better use of the data for data mining and decision support, it should be changed into high-quality data. Therefore, data cleaning, will be used to clean up the dirty data before using the data. Consequently, the following section will show how to remove the dirty data from the above six selected tables: Step1. Remove dirty data. Customer's table has 2384 customers’ information, but there are 141562 recorders in the card trading log table that represent the transaction information of 2180 customers. Thus, 204(2384-2180) customers’ information will be removed from customer's table. The recorders of transactions not happening in the current year also should be deleted from the balance history table. There are 2689 pieces of information about the overdraft records of the 2180 customers whose transaction information has been saved in the card trading log table. So the 75(2764-2689) recorders will be abandoned. Step 2. Handle null values. Because some fields are not required, there are a few null values in the original database. We deal with the null values, mainly by the following means. For family population field, null values have been replaced with 4 because the statistics of the recorders in the customers indicate that the larger proportion of families has four members. The blanks in the transaction money field have been replaced by average of transaction money. For age field and occupation field, the null value will be replaced by the value of the center point of each class got by K-means algorithm. For monthly income field, the value can be evaluated based on the salary information of the cardholders if their salary is distributed by the bank. Other null values are replaced by the value of the center point of each class. Step 3. Rectify false data. Do search and analysis of the data in the data sheet through a query analyzer, to see if there exists any value, which cannot match the corresponding fields. The corrections of false data are difficult to do, however, in most cases, the estimation of one certain field’s value can lay a foundation for the measure of another field’s value. For example, we can estimate one’s credit standing according to his resume and occupation. Step 4. Removes to duplicate data. Make use of the key word ‘distinct’ in ORACLE to clear the recurring data in the process of data process. In order to reduce and avoid the redundancy and inconsistency in result data set, we should make the mutual-related different data source got together. Join the former six sheets and temporary data sheet together with card-id and account number, from which select such 20 fields as card_id, sex, age, family_population, occupation and so on. Accordingly, the above-mentioned fields constitute a data set, and in which there exist many fields of continuous attribute. In


1101


order to improve the learning speed of the neural network and avoid network overload, the evaluation elements, such as sex, age, marital status, educational attainment, occupation, transaction amount, month income, and so on, need to be zoomed in proportion to ensure the values within the uniform specified scope before we input them into the neural network. In other words, the attribute values will be in the section of [0, 1] by normalization. For example, the value v of the attribute A will be mapped to v’ by the following formula: v′ = ( v − min A ) / ( max A − min A )

V.2.

(10)

VI.

VI.1. The Choice of Experimental Data In our experiment, we pick out a certain type of customers. There are 237 credit customers in this type. We choose their transaction data to detect credit card fraud from January 5 to March 30 in 2003, including 18581 normal trading and 130 risk trading. The total number of data is divided into two groups, named training set and testing set. We select 2813 transactions from normal trading and 87 transactions from risk trading as the training set, except 43 risk transactions and 5375 normal transactions as the testing set.

Customer Segmentation

The growth of data stocks leads to new challenges for building up real-time information systems with the purpose of prediction. The predictions made by Chaos-BP prediction model are not satisfactory. Chaos-BP prediction will become ineffective when the sample point is too large. So we conducted dynamic customer segmentation model based on customers’ behavior by combining k-means algorithm with cloud model to promote the accuracy of predicted results [16]. V.3.

TABLE I THE CHOICE OF EXPERIMENTAL DATA Risk-taking data Training set 1~87

denoted as X = ( X1 "X16 )

T


Training set 1~2813

Testing set 2814~8188

The chaos-BP neural network model is used to training set, and testing set respectively. Test results obtained in Table II. TABLE II THE RESULTS OF CHAOS-BP NEURAL NETWORK MODEL Training set Value 0 1

T

Step 1. Determine the delay time τ and the minimum embedding dimension with C-C method, and then reconstruct the phase space of chaotic time series to constitute the study sample and teachers value. Step 2. Determine the structure of neural network. The embedding dimension is used as the number of input layer, and the number of output layer is 1. Step 3. Network learning and training. N-1 points and their corresponding outputs are regarded as samples for training until the error is controlled within the allowable range or the number of iterations reaches its intended value. Step 4. Model predictions. While N-phase is input to the chaos-BP model, the output is the just predictive value.

Testing set 88~130

VI.2. Comparison and Evaluation of Experimental Results

. Vector Y = ( y1 ,y2 ,y3 )

represents the three nodes of the hidden layer and O1 denotes the node of the output layer:

Normal data

Table I shows that the number of training samples in training set is 2890, including 87 from risk trading and 2813 from normal trading, respectively, the number of testing samples is 5418.

Prediction Steps

Through a large number of numerical experiments, we find that the prediction effect will be better when the number of input layer is equal to the embedding dimension m of state space reconstruction of chaotic time series [12]. In this paper, according to the characteristics of credit card consumption data, the phase-space should be reconstructed by the above method. The number of input layer is 13 in the study. So we design a three-layer neural network, which having thirteen nodes in the input layer

Experiment

Forecast 0 1 2797 16 17 70

Detection rate TP/FN% FN=0.57 TP=80.46

Testing set Forecast 0 1 5357 18 9 34


In order to compare the effect of predict, a simple BP neural network model is also used to detect credit fraud. For BP neural network model, we selected 18 neurons in input layer, corresponding to 18 factors in transaction risk detection. The number of output layer is also 1. Test results obtained in Table III. TABLE III THE RESULTS OF THE BP NEURAL NETWORK MODEL Training set Value

0 1

Forecast 0 2795 21

1 18 66


Testing set Forecast 0 5367 10

1 148 33


Detection performance is mainly determined by two factors, true positive prediction rate (TP) and false negative error rate (FN): International Review on Computers and Software, Vol. 6, N. 6

1102


TP=the number of correctly identified risk transactions /total number of risk transactions. - FN= the number of normal transactions detected as the high-risk transactions /total number of normal transactions. The results show that prediction accuracy of the proposed model is higher than neural network prediction method, and the correct detection rate used by Chaos-BP algorithm after customer segmentation has increased slightly, while error detection rate has decreased. -

VII.

Conclusion

In the paper, a dynamic customer segmentation model based on cloud theory has been primarily implemented to promote the accuracy of predicted results. Then a model based on neural network model and chaos theory has been built up to detect risk-trading for each kind of clients. Experiments show that true positive prediction rate of the proposed model is 3% higher than BP neural network prediction method.

[15] Linghui Zhai, Shaoping MA, and Huanling Tang, Data Preprocessing of Classification Mining of Credit Card, Computer Engineering, no. 6, pp. 195-197, 2003. [16] Haining Tan and Juanjuan Xu, Study on Customer Segmentation Theory Based on Cloud Model: A Case of Telecom Industry in China, Technology Economics, vol. 28, n. 4, 2009, pp. 54-56.

Authors’ information School of Information and Engineering, Henan Institute of Science and Technology, Xinxiang 453003, Henan, China. Yanli Zhu was born in 1976, she received the B.S degree form Computer Henan normal university in 1999 and her M.S degree from Sichuan Normal University in 2007. She is a faculty member in the School of Information Engineer, Henan Institute of Science and Technology. Her current research interests are data mining and image processing. She is a membership of China Computer Federation.

References [1]

[2] [3] [4]

[5] [6]

[7]

[8]

[9]

[10]

[11] [12]

[13]

[14]

Peican Jiang, A study of credit card risk transactions’ real-time early warning system, South China University of Technology, 2004. Hua Yan and Menliang Hu, System design of avoiding credit card fraud, Micro-computer information, December 2006. R. J. Bolton, D. J. Hand, Statistical Fraud Detection: A Review, Statistical science, vol. 17, n. 3, pp. 235-237, 2002. F. S. Maes, K. Tuyls, B. Vanschoenwinkel, B. Manderick, Credit Card Fraud Detection Using Bayesian and Neural Networks, Proceedings of Neuro Fuzzy, Havana, Cuba, 2002. Dawei Li, Haiying Ma, and Zhou, Credit card fraud detection and early warning of risks，Investment Research, n. 1, pp. 23-26, 2007. Yongkuan Qin, Shengxiang Huang, and Qing Zhao, A predition research of deformation with chaos theory and artificial neural network, Journal of Geomatics, vol. 34, n. 1, pp. 40-42, 2009. W. Jonathan, P. K. Agarwal, J. Baker, Modeling and prediction of chaotic systems with artificial neural networks, International Journal for Numerical Methods in Fluids, vol. 63, n. 8, pp. 989-1004, 2010. Smith, A. Matthew and Kaplan, Lev, Integrating random matrix theory predictions with short-time dynamical effects in chaotic systems, Physical Review E-Statistical, Nonlinear, and Soft Matter Physics, vol. 82, n. 1, 2010 O. I. Antipov, V. A. Neganov, Neural network prediction and fractal analysis of the chaotic processes in discrete nonlinear systems, Doklady Physics, vol. 56, n. 1, pp.7-9, 2011. Bin Wang, Jun Tang, and Yeqing Tao, A method for outlier time series detection based on one-step chaotic prediction, Engineering Journal of Wuhan University, vol. 43, n. 2, pp. 265-268, 2011. Deyi Li and Changyu Liu, Study on the Universality of the Normal Cloud Model,” Engineering Science, vol.6, n. 8, 2004, pp. 28-33. Lu Zhen-bo, CAI Zhi-ming, and JIANG Ke-yu, Determination of Embedding Parameters for Phase Space Reconstruction Based on Improved C- C Method, Journal of System Simulation, vol. 19, n.11, pp. 2527-2529, 2007. Takens F, “On the numerical determination of the dimension of an attractor, dynamical system and turbulence, Lecture Notes in Mathematics, Berlin: Springer-Verlag, 1981, pp. 230-241. Ruihua Lv, Weiya Wang, Research on the Mixed Forecasting Methods of Chaotic Time Series, China Soft Science magazine, n.2, pp. 150-154, 2006.



1103


Buyer-Seller Watermarking Protocol without Trust Third Party Lv Bin1, Fei Long2

Abstract – Piracy becomes increasingly rampant in recent years and has caused a great loss. As a prominent technology, digital watermarking has been developed to combat piracy. In a real watermark application, the buyer and the seller follows a secure buyer-seller watermarking protocol, can enable a seller to successfully identify a traitor from a pirated copy, while preventing the seller from framing an innocent buyer. A trust third party (TTP) is required in some known watermarking protocols, which may affect the security as the seller or the buyer may collude with the introduced TTP to cheat the other. A buyer-seller watermarking protocol is proposed in this paper, where no TTP is required without reducing the security. In the proposed watermarking protocol, homomorphic cryptosystem and bit commitment are used to achieve the asymmetric property and solve the disputation. Compared to the previous protocols, the proposed protocol has good security and can be easily implemented as no TTP is introduced. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Buyer-Seller Watermarking Protocol, Copyright Protection, Homomorphic Encryption, Bit-Commitment

I.

Introduction

With the rapid advancement in networking and computer technology, e-commerce has been a common business way, especially for some digital products such as softwares and different kinds of multimedia copies. The growing e-commerce in digital industry make people buy and use all kinds of digital products easily. Unfortunately, piracy becomes increasingly rampant as customers can easily duplicate and redistribute the received digital product to a large audience without quality loss. To protect the intellectual property rights of the digital industries and promote digital products related services, ensuring the proper distribution and usage of digital products has become increasingly critical, especially considering the ease of digital data. Encryption is a traditional and effective data protection technology, which has played an important role in digital copyright protection area. However, encryption only can provide digital data with the desired security during transmission by preventing them from unauthorized accessing. Once a piece of digital data is decrypted, this protection is disappeared as the dishonest customer can redistribute it arbitrarily. As a perfect complement of encryption, information hiding has developed to combat piracy by embedding a copyright notice (watermark) into the original digital products [1]. Once an unauthorized copy is found somewhere, the owner can prove the ownership by extracting the watermark. In a real watermark application, each participant such as the buyer and the seller follows a secure buyer-seller watermarking protocol that combines digital watermarking and cryptography [2], [3]. Manuscript received and revised October 2011, accepted November 2011

1104

A secure and fair buyer-seller watermarking protocol should enable a seller to successfully identify a traitor from a pirated copy, while preventing the seller from framing an innocent buyer. Buyer-seller watermarking protocol has been a hot research in both e-commerce and information security fields. A number of buyer-seller watermarking protocols have been proposed in these years [4], [5], [6]. In some known watermarking protocols, a TTP such as the watermark certificate authority was introduced to guarantee the fairness of both parties [4], [5]. However, the introduced TTP may decrease the security, as the seller (or the buyer) may collude with the TTP to cheat the buyer (or the seller). This problem usually is called conspiracy. If the introduced third party is untrustworthy, the conspiracy problem has to be considered. In addition, the introduced TTP increases the complexity of implementation, as the complexity of three-party interaction is more than the two-party interaction. Moreover, two-party transaction is more similar to the transaction of real world.

II.

Related Works

The first buyer-seller watermarking protocol was proposed by Qian and Nahrstedt to solve the problem that a dishonest seller may frame an innocent buyer [2]. In this protocol, the buyer sent the seller an encrypted watermark that will be embedded by the seller. Since only the buyer knew the decryption key, he can prove his ownership of the digital product that he bought. Memon and Wong improved the protocol proposed by Qian and Nahrstedt to


Lv Bin, Fei Long

solve the problem that a malicious seller can illegally redistribute an innocent buyer's copy and accuse him of pirating [4]. In [5], the authors enhanced Memon and Wong's scheme to solve problem that binding a chosen watermark to a given digital product by introducing a common agreement between the seller and the buyer. In [6], the authors have proposed an improved buyer-seller watermarking protocol to solve the problem that a double watermarking insertion may decrease the quality of final watermarked copy. The homomorphic public key cryptosystem has been explored as a basic tool to achieve some of the well known watermarking protocols [4], [5]. In these protocols, the encrypted watermarks can be directly embedded into the encrypted digital product without prior decryption. The additive homomorphic cryptosystems are most popular in the partially homomorphic cryptosystems. An encryption scheme is said to be homomorphic if for any given public encryption key pk , the encryption function E satisfies: E pk ( x + y ) = E pk ( x ) E pk ( y ) ,

(1)

where x and y are the plaintexts. A general and more detailed introduction to the homomorphic public-key cryptosystems can be found in [4]. Bit commitment is an important cryptographic primitive in modern cryptography and widely used to design secure protocols [7]. A more detailed introduction of bit commitment can be found in [8] and [9]. Generally speaking, a bit commitment scheme is a two-party (the committer and the receiver) protocol, which includes two stages, which are committing stage and revealing stage. Through this protocol, the committer can be blinded with some particular information. In the committing stage, the committer commits a piece of information to the receiver, but the receiver cannot infer which the committed information is. In the revealing stage, the committer can prove to the receiver revealed information is same to the committed information. The public key cryptography is the common way to design bit-commitment schemes.

III. Proposed Watermarking Protocol The basic idea of the proposed secure buyer-seller watermarking protocol in this section can be depicted as follows. For fairness, the final watermark consists of two parts, one is from the seller and the other is from the buyer. In order to solve the potential disputation and achieve the requirement of non-reputation, both the buyer and the seller should make a commitment for his information. The proposed watermarking protocol is performed between the seller Alice and the buyer Bob. The proposed protocol has three sub-protocols, which are registration protocol, watermarking protocol, copyright identification and dispute resolution protocol.


III.1. Registration Protocol Privacy protection in e-commerce transaction is very important, as the malicious seller may make use of the information to benefit from reselling these data to other parties or making criminal actions. Similar to the previous protocols [5], if Bob wants to stay anonymous during the transactions, he can request to the certification authority (CA) for an anonymous certificate in advance, via the registration sub-protocol presented in this subsection. Generally speaking, an anonymous certificate is a normal digital certificate except that the content of its subject field is a pseudonym rather than the real identity of the applicant. By assigning the anonymous certificate to Bob, CA is responsible for binding this particular anonymous certificate to the buyer Bob. CA also guarantees that the binding is not revealed to any other party unless requested by judge when Bob is proven to have committed piracy. To apply for an anonymous certificate, Bob first randomly selects a public-private key pair ( pkb ,skb ) and sends pkb to CA. When CA receives pkb from Bob, it generates an anonymous certificate CertCA ( pkb ) and

sends it to Bob. Here, similar to [5], we let the public key pkb be the pseudonym associated with the anonymous certificate assigned to Bob, as mentioned in the standard of Simple Public-Key Infrastructure, which allows for the use of pseudonymous public keys. III.2. Watermarking Protocol To carry out a transaction, Bob and Alice follow the watermarking protocol described in this subsection: Step 1. To buy a copy of digital product titled by X from the seller Alice, Bob first negotiates with Alice to set up a common agreement, ARG, which explicitly states the rights and obligations of both parties, and specifies the digital content of interest. ARG uniquely binds this particular transaction to X and can be regarded as a purchase order. Note that Bob may use his pseudonym in the negotiation to keep his identity unexposed. In addition, ARG should point the way how to generate a watermarked copy and the trading guide. The detailed generation process of the watermarked copy will be introduced step by step. Step 2. To buy a digital product X , Bob generates a watermark wb guide pointed by ARG, which is a vector of integers with the length of L . The arrangement of wb is similar to [5], where the first l1 elements are zeros and the last l2 = L − l1 elements are integers of non-zeros. Then Bob sends his public key pkb and his commitment of pkb denoted by C [ pkb ] , the encrypted wb denoted

by E pkb ( wb ) and the commitment of wb denoted by C [ wb ] to Alice.


1105

Lv Bin, Fei Long

Step 3. Upon receiving the request from Bob, Alice generates a watermark wa according to the guide pointed by ARG. The watermark wa is also a vector of integers with the length of L . Differing from wb , the first l1 elements of wa are integers of non-zero and the last l2 = L − l1 elements are zeros. Then Alice encrypts wa with pkb , and then she generates the composited watermark and embeds it into the digital product X requested by Bob by utilizing a homomorphic public-key cryptosystem. Finally, Alice sends the encrypted watermarked copy denoted by E pkb ( X w ) and the corresponding commitment to Bob. Step 4. When Bob receive E pkb ( X w ) , he can decrypt it with his private decryption key skb to get the finally marked copy. III.3. Copyright Identification and Dispute Resolution Protocol Once an unauthorized (pirated) copy X of a certain digital product X owned by the seller Alice is found somewhere. Copyright identification protocol depicted in this subsection can be used to track the identity of the responsible distributor, who was the buyer in some earlier transaction. The copyright violator identification protocol is executed by the seller Alice. To track the leak source of the pirated copy, Alice first encrypts the detected copy X with a list of m public encryption keys pki ( i = 1, 2 ," m ) that is stored in her local transaction database. Here, m is the number of buyers. Then Alice can locate the corresponding buyer who made a piracy by comparing

( ) ( i = 1,2," ,m ) .

E pkb ( X w ) with each E pki X

In case Bob denies that the unauthorized copy X has originated from his version, Alice collects the associated information and sends them to Judge for arbitration. In the proposed watermarking protocol, the most important information provided by both parties for arbitration is the corresponding commitments. If the suspected buyer Bob wants to prove his innocent, he only needs to reveal his commitments. If the corresponding commitments of Bob are consistent with the detected results, then Bob is considered to be a guilty buyer.

IV.

Discussion

In this section, we analyze the proposed buyer-seller watermarking protocol from the security and practicality. It has to be first pointed out here that the security of the proposed watermarking protocol related to the security of the used cryptographic primitive, which is the adopted homomorphic encryption cryptosystem. The proposed watermarking protocol is secure for the buyer Bob. First, the seller Alice doesn’t know Bob’s


watermark wb , and hence he cannot get to know the finally composited watermark embedded in the digital product X bought by Bob. In this case, Alice cannot frame Bob by forging Bob’s copy. In addition, since Alice does not know Bob’s encryption key pkb , Alice cannot redistribute the digital product bought by Bob directly. The proposed watermarking protocol is also secure for the seller Alice. First, once an unauthorized copy of digital product X is found somewhere, Alice can trace the suspected buyer effectively by extracting the embedded watermark, especially for the watermark wa . In addition, because Bob only knows his watermark wb , not the composited watermark, which is generated from wa and wb . Therefore, it is infeasible for Bob to remove his watermark wb from the watermarked copy. Moreover, Bob cannot claim that the copy was created by Alice. Because only Bob knows the private decryption key skb and wb , no one can forge Bob’s copy. In other words, Bob cannot deny his responsibility for a copyright violation caused by him. Since no any TTP is introduced in the watermarking protocol, the problem that the seller or the buyer colludes with the TTP to cheat the seller or the buyer is avoided, which is secure and fair for both parties. In addition, the protocol is easier to implement and more similar to real world. The proposed buyer-seller watermark protocol avoids a double watermark insertion, and then the final quality of the watermarked content is higher and the result of watermark detection is more accurate. The homomorphic public-key encryption algorithms are based on the algebraic property of integers. However, the digital products are usually expressed by real numbers. This is a limitation of homomorphic public-key encryption. The proposed watermarking protocol has taken this limitation of the homomorphic public-key encryption into account, which will make the protocol more practical.

V.

Conclusion

In this paper, a secure and practical buyer-seller watermarking protocol has been proposed based on some known protocols. Compared to previous watermarking protocols, the proposed protocol has some important improvements. An important feature of the proposed protocol is that there is no need to require a TTP in the protocol without decreasing the security of the whole protocol. In addition, the issues of double watermarking insertion and the limitation of homomorphic encryption cryptosystem are considered jointly. The non-reputation property of the bit-commitment can solve the potential disputation well.


1106

Lv Bin, Fei Long

Acknowledgements


This paper is supported by the project of National Natural Science Foundation of China under grant No. 71073118.

References [1] [2]

[3]

[4]

[5]

[6]

[7]

[8] [9]

M. Barni, F. Bartolini, Data hiding for fighting piracy, IEEE Signal Processing Magazine, vol. 21, n. 2, pp. 28-39, 2004. L. Qiao, K. Nahrstedt, Watermarking schemes and protocols for protecting rightful ownership and customer's rights, Journal of Visual Communication and Image Representation, vol. 9, n. 9, pp. 194-210 , 1998. F. Frattolill, Watermarking Protocol for Web Context, IEEE Transactions on Information Forensics and Security, vol. 2, n. 3, pp. 350-363, 2007. N. Memon, P. W. Wong, A buyer-seller watermarking protocol, IEEE Transactions on Image Processing, vol. 10, n. 4, pp. 643-649, 2001. C. L. Lei, P. L. Yu, P. L. Tsai, M. H. Chan, An efficient and anonymous buyer-seller watermarking protocol, IEEE Transactions on Image Processing, vol. 13, n. 12, 1618-1626, 2004. Defa Hu, Qiaoliang Li, A practical and secure buyer-seller watermarking protocol, Journal of Digital Information Management, vol. 9, n. 1, pp. 43-47, 2011. Caimu Tang, D. O. Wu, Efficient multi-party digital signature using adaptive secret sharing for low-power devices in wireless networks, IEEE Transactions on Wireless Communications, vol. 8, n. 2, pp. 882-889, 2009. M. Naor, Bit Commitment using pseudorandomness, Journal of Cryptology, vol. 4, n. 2, pp. 151-158, 1991. S. Winkler, J. Wullschleger, S. Wolf, Bit commitment from nonsignaling correlations, IEEE Transactions on Information Theory, vol. 57, n. 3, pp. 1770-1779, 2011.


1,2

Center for Studies of Information Resources, Wuhan University, Wuhan, Hubei, China, 430072. Lv Bin was born in Chaohu, China, in 1980. He finished his Master Degree of information science from the Department of Information Science, Anhui University, in 2009. He has published some books and articles such as “Usability Evaluation of E-Commerce Website with the Analytic Hierarchy Process” (ICCNT 2011). His research interests are competitive intelligence, information resources management. Mr. Lv currently is a candidate for Doctor Degree of information science from Wuhan University.. Fei Long was born in Yueyang, China, in 1983. He finished his Master Degree of computer science from the Department of Software, Hunan University, in 2008. His research interests are business intelligence, recommender system. Mr. Long currently is a candidate for Doctor Degree of information science from Wuhan University.


1107


Ontology-Based Oracle Bone Inscriptions Machine Translation Jing Xiong1,2, Lei Guo1, Yongge Liu1,2, Qingsheng Li1

Abstract – Oracle Bone Inscriptions (OBI) refers to incised ancient Chinese characters found on oracle bones, which are animal bones or turtle shells used in divination in Bronze Age China. The vast majority record the pyromaniac divinations of the royal house of the late Shang dynasty at the capital of Yin. OBI is the earliest and complete system ancient Chinese characters. It has very important research merit. With the development of computer and information technology, digitization processing of OBI becomes an important aspect in current digitization processing of ancient Chinese font. But in the digital processing, the burden of OBI experts is heavy and the knowledge-sharing degree of OBI is low. In order to resolve these problems, proposed a solution of machine translation based on ontology. Firstly, the model and processes of OBI ontology construction are analyzed. Secondly, the flow of ontology-based machine translation and its key technologies are introduced. Finally, further research work is prospected. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Oracle Bone Inscriptions, Ontology, Machine Translation, Semantic Similarity, Corpus

I.

Introduction

OBI has more than 3,500 years’ history. It records a wide range of social activities of the Shang Dynasty royal divination. Because of its very rich historical content, it has important research merit [1]. But it is very difficult to identify and collate OBI. Using information technology to study OBI is a very important way. However, OBI has high study difficulties. Currently, OBI can be identified and translated by only a few persons, and OBI professional training needs one or two decade or even longer [1]. Many OBI experts have pointed out it is an interesting and meaningful work to translate OBI into modern Chinese [2]. OBI belongs to Chinese ancient characters. There have been some machine translation researches about ancient characters. Reference [3] introduced an example-based ancient prose machine translation system named EBMTAC in order to avoid complex deep syntax and semantic analysis. Reference [4] studied the automatic rhythm selection of Chinese poetry in English translation by using statistical models, and summarizes some important factors which impacted the English rhythm. OBI language has many similar features to modern Chinese ， the translation between them just like the one between modern Chinese and minority languages. Reference [5] used frequency statistics and frequency distribution statistics as collection principles to design an electronic dictionary for Uyghur-Chinese machine translation ， describing the dictionary constructor, sort principle, the index table, data structures of attribute library and dictionary information retrieval methods.


1108

Reference [6] designed and implemented an EBMT Chinese-Mongolian system. It studied the instance search algorithm and methods of phrases fragmenting, matching and combing. However, these studies lack the semantic analysis of text, and they are hard to deal with polysemy, synonyms and other issues. It records OBI experts’ identification and translation relies on their long-term academic study and experience. This experience knowledge is only stored in the minds of experts, and effective sharing of knowledge cannot be achieved. It is a major issue for machine to understand and sharing the knowledge in OBI information processing. Ontology as a shared and clear conceptual model of formalization can provide a solution to this problem. It also supports semantic analysis. Began the 1990s, researchers had used ontology technologies to establish natural language understanding systems and machine translation systems, such as PENMAN, Mikrokosmos, ONTOS and SENSUS. In [7] an ontology-based Japanese-English dictionary is introduced. HowNet [8] and HNC [9] are also ontology related. Both [10] and [11] have studied English-Chinese or Chinese-English machine translation based on ontology. The aim of this paper is to improve OBI knowledge sharing and reduce the study difficulties of OBI. So we proposed a machine translation strategy to assist information researchers processing OBI and decrease the dependence of experts. In our scheme, we based on electronic dictionary and use ontology as middle language to implement OBI computer-aided translation. Compared to existing scheme, our scheme has good


Jing Xiong, Lei Guo, Yongge Liu, Qingsheng Li

semantic ability in dealing with multi-category words and synonym problem in OBI.

II.

OBI Ontology Building Model

OBI and modern Chinese are closely linked. Many Chinese characters coining methods and grammars are survived into modern Chinese. On one hand, we can reference and reuse of modern Chinese classification system, making the top-down detailed description. On the other hand, with the OBI research progresses, more and more knowledge and rules are found, we can summarize the existing results and classify them, for the gradual bottom-up abstract. Based on this point, we proposed ontology building model as shown in Fig. 1.

HowNet is an on-line common-sense knowledge base unveiling inter-conceptual relations and inter-attribute relations of concepts as connoting in lexicons of the Chinese and their English equivalents. The units for manipulation and description in HowNet are thing (sub-divided into physical and mental), Part, Attribute, Time, Space, Attribute-value and Event [8]. So, it can provide a reference and learning for us to extract ontology concepts. OBI ontology construction process is shown in Fig. 2.

Fig. 1. OBI ontology building model Fig. 2. OBI ontology construction process

The model shown in Fig. 1 is called "two-way active hinge method". In the vertical direction, using top-down and bottom-up combination approach. Top-down means planning a top-level ontology concept category first, and then concretizes the ontology by expanding subcategory; bottom-up means extracting the common characteristics among the domain terms and classify them into one category, and then keeps on abstracting. Intersection in the two directions is called "twist point". It is an appropriate level and has formed an ontology prototype. "Twist point" is active, because it is hard to achieve an appropriate level one time. Ontology construction process is a dynamic process that requires iterative, gradually improved.

IV.

OBI Machine Translation Process

OBI machine translation is based on OBI dictionary and ontology. First, the input OBI annotation sentences will be segmented into words. Second, using the OBI dictionary, the OBI words will be translated into corresponding modern Chinese words. Third, OBI ontology provides semantic concepts for the semantic analysis of the entered words, helping realize semantic representation and eliminate the ambiguity. Last, combine the word segments based on syntax rule library and output the target modern Chinese sentences. The machine translation process is shown in Fig. 3.

III. OBI Ontology Construction Ontology as the intermediate language between the source language and the target language needs to meet certain requirements. It is not another expression of OBI dictionary, and not a collection of entries. The main function of the ontology here is to express the semantics of the words and sentences in the source language and the target language. It is a shared conceptual model, so it is independent of the specific language form. When building OBI ontology, it needs to gain the common concepts and the concept relationships between OBI and modern Chinese. Therefore, we should build an abstract model based on the two languages. OBI and modern Chinese are closely linked. In modern Chinese, many elements extend from OBI such as coining Chinese characters and Chinese grammar. So we can reference and reuse the modern Chinese classification system, implementing the top-down progressive refinement.


Fig. 3. OBI ontology building model

IV.1. OBI Word Segmentation Word segmentation is a very important step for OBI machine translation. The process of our OBI word


1109


segmentation method is as follows. First, extract the corresponding inscriptions text from OBI corpus to. Second, finish the rough word segmentation based on dictionary. Third, parse the syntax through bottom-up method. Fourth, handle with the ambiguous words and unlisted words using syntactic rule base, checking the unknown words and recall the valid ones into the OBI dictionary at the same time. Finally, optimize the result of word segmentation. The OBI word segmentation flow chart is shown in Fig. 4.

The semantic weight values please refer to [12].The semantic similarity of one concept is in inverse proportion to the semantic distance. When the semantic distance is 0, the semantic similarity is 1, and when the semantic distance is infinite the semantic similarity is 0. The semantic similarity can be simplified as:

(

)

Sim Ci ,C j =

(

1

(2)

)

Dist Ci ,C j + 1

Calculating all the semantic similarities of the multi-category words and the word which has the largest value is the best one.

V.


This section shows some experimental results to examine the performance of the proposed scheme. We choose 30 OBI sentences as test samples. They are all simple sentences including subject-predicate (S-P) sentences and non-subject-predicate (N-S-P) sentences. Table I shows the precision, recall rates and F values of our word segmentation method. Fig. 4. OBI word segmentation process

TABLE I COMPARE OF PRECISION, RECALL RATE AND F VALUE

IV.2. Concept Mapping

Segmentation tool

Precision

Recall

F-value

There are many multi-category words in Chinese, and we should determine which one is the best choice for the target sentence. Concept semantic similarity can help us to

ICTCLAS

66.7%

71.1%

68.8%

our method

97.5%

97.6%

97.6%

(

choose the best one. To define Dist Ci ,C j

)

as the

semantic distance between two ontology concept Ci and C j . The following equation calculates the semantic distance:

(

) ∑ω

Dist Ci ,C j =

n

k =1

ek

+

NCi + NC j NCi + NC j + 2 × N LCA

× ε (1)

where: n

-

∑ ωe k =1

k

is the total weighted distances of the edges

within the shortest path which links to concept node Ci and C j . -

NCi is the weighted distance between Ci and the

lowest common ancestor note Ca of Ci and C j , so does NC j . -

N LCA is the weighted distance between Ca and the root node. ε is a constant determined by weighted coefficient.


Institute of Computing Technology, Chinese Lexical Analysis System (ICTCLAS) is a very popular analysis system for modern Chinese. However it is not good at dealing with OBI. Our word segmentation method fits the OBI field, and its efficiency is much higher than other methods which oriented modern Chinese word segmentation. The main reason is that the method based on OBI special dictionary and it follows the syntax rules of OBI. Higher recall rate because the vast majority words in the OBI are one-character words. It was also found that word segmentation accuracy rate increase with the augment of OBI corpus. Table II shows the average accuracies of ontology-based OBI machine translation. The conclusion is drawn by OBI experts artificially. TABLE II AVERAGE ACCURACIES OF DIFFERENT SENTENCE PATTERNS Sentence Pattern (AV) + S + V S+V (AV) + S + AV + V S + V + CO AV + V + O (AV) + noun

Sentence Type S-P S-P S-P S-P N-S-P N-S-P

Sentence Number 7 10 4 5 2 2

Average Accuracy 62.2% 80.3% 56.5% 63.4% 65.0% 72.1%


1110


In our experimental sentences, S means subject, V means predicate head, AV means adverbial, CO means complement and O means object. The round bracket such as (AV) in the sentence indicates the element within the bracket can appear or not. We can see that the input sentences which have simple structure have higher accuracy value than the complex ones. Most of the segmented words can find corresponding modern Chinese in OBI dictionary and the performance of translation is high depend on it, ontology helps it to finish word sense disambiguation. If the word orders of the two languages are consistent, it causes high quality translation. Because there is no other OBI machine translation system, we have no contrast experiment.

VI.

[8]

Liao Jian, Leng Jing, Li Yanyan, Huang Ronghuai, Research on HowNet based on formal concept analysis and concept similarity, Application Research of Computers, vol. 24, no. 11, pp. 32-36, 2007. [9] Cheng Xianyi, Zhu Qian, Han Fei, Semantic Chunk of Question Sentence Analysis Based on HNC and Description Logics, Journal of Guangxi Normal University: Natural Science Edition, vol. 28, no. 3, pp. 131-134, 2010. [10] Wang Xiaojie, Zhong Yixin, Using Ontology in an English-Chinese Machine Translation System, Journal of Chinese Information Processing, vol. 14, n. 5, pp. 8-15, 2000. [11] Shi Chongde, Wang Huilin, A Research on the Ontology-based Chinese-English Machine Translation, Library and Information Service, vol. 50, n. 9, pp. 14-17, 2006. [12] Xiong Jing, Wang Ai-Min, Xu Jianliang, Information retrieval optimization strategy based on domain ontology, Computer Engineering and Design, vol. 32, n. 8, pp. 2695-2699, 2011.


Conclusion 1

This paper investigated the ontology-based machine translation technology for OBI. The purpose is to fully share the existing knowledge of OBI, reduce the burden of OBI experts and lower the difficulties of OBI study. It analyzed the key technologies about OBI ontology construction. Because of the scale and semantic limit of ontology library, experiments of complex sentences have not yet fully. In the future, we will expand and perfect the OBI ontology and do some experiments to further verify the effectiveness of our methods. The evaluation mechanisms for OBI machine translation is also a key research issue.

School of Information and Engineering, Anyang Normal University, Anyang 455002, Henan, China. 2

Oracle Bone Inscriptions Information Processing Laboratory, Anyang 455002, Henan, China. Jing Xiong received his bachelor’s degree from Jinan University in the year of 2001. After that, he got his master’s degree from Ocean University of China in 2006. And he received his Ph.D. from Ocean University of China in 2010. Now, he is a lecturer in the School of Computer and Information Technology, Anyang Normal University. His major research interest is in semantic web, knowledge representation and reasoning, ontology theory and application.

Acknowledgements This work was supported by National Natural Science Foundation of China (60875081) and National Natural Science Foundation of China (60973051).

References [1] [2]

[3]

[4]

[5]

[6]

[7]

Jiang Ming-hu, Natural Language Processing, BeiJing: Higher Education Press, 2006. Wang Yu-xin, Yang Sheng-nan, NIE Yuhai, Oracle Bone Inscriptions pithiness explanation, Yunnan People's Publishing House, 2004. Wang Shuang, Xiong Delan and Wang Xiaoxia, The Research and Implemention of Example-based Machine Translation of Ancient Chinese, Journal of Xuchang University, vol. 28, n. 5, pp. 88-91, 2009. Fang Miao, Jiang Xin, Zhao Qi and Jiang Yi, Automatic Choosing of English Rhymes in Translation of Chinese Ancient Poems, Proceedings of the 2nd International Symposium on Knowledge Acquisition and Modeling (Page: 434 Year of Publication: 2009 ISBN: 978-0-7695-3888-4). Alifu.Kuerban Abulimiti, Abudureyimu Tuergen and Yibulayin, Design of a Dictionary for Uighur-Chinese Machine Translation System, Computer Engineering and Applications, vol. 42, n. 20, pp. 76-78, 2006. Hou Hongxu, Liu Qun, Nasun Urt, Example Based Chinese -Mongolian Machine Translation, Journal of Chinese Information Processing, vol. 21, n. 4, pp. 65-72, 2007. A. Okumura, E. Hovy, Building Japanese-English dictionary based on ontology for machine translation, Association for Computational Linguistics, 1994.



1111


A Method for Mining Association Rules Based on Cloud Computing Fei Long, Yufeng Zhang, Lv Bin

Abstract – Apriori algorithm is one of the most classical algorithm to extract association rules. However, the traditional Apriori algorithm is only suitable for analyzing and mining the centralized data. Through the analysis of the Apriori algorithm which had some defects, and the advantages of cloud computing platform demonstrated in large cluster. To improve the traditional Apriori algorithm, in this paper, a method for mining Association Rules Based on Cloud Computing is proposed. The new method used HDFS to store data and is well adapted to the Hadoop's MapReduce computing model. It inherits the MapReduce scalability to huge datasets and to thousands of processing nodes. Experimental results show that it is very efficient compared with traditional Association Rules algorithm and have a good speedup when deals with massive data. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Association Rule, Cloud Computing, Apriori Algorithm, Hadoop

I.

Introduction

As massive data sets become increasingly available, people are facing the problem of how to effectively process and understand these data. Data mining as an efficient way to extract the knowledge has draw attention from both the research and industrial communities. Conventional data mining algorithms are developed with the assumption that data is memory resident, making them unable to cope with the exponentially increasing size of data sets. Association rule is one of the important themes and essential to data mining which can find out the relationship between item sets in the database. We can use the interesting association relationships which are extracted among huge amounts of data. As association rules widely used, it needs to study many problems, one of which is the generally larger and multi-dimensional datasets, and the rapid growth of the data. The discovery of association rule is a direct mass-oriented database system that often has hundreds of properties and millions of records, contains a complex relationship between data tables, and remains a time-consuming process. This will inevitably lead to a great surge in search of dimension and space. Apriori algorithm finds all the frequent itemsets by scanning the database time after time, and it will consume a lot of time and memory space when scanning the database with mass data, which will become the bottleneck of Apriori algorithm. Advances in computing and networking technologies have resulted in many distributed computing environments, the use of parallel and distributed systems has gained significance. Single processor’s memory and CPU resources are very limited, which makes the algorithm performance inefficient.


1112

The parallel association rule algorithm with high performance is best choice. Currently proposed parallel algorithms are Count Distribution, Candidate Distribution and Data Distribution [1]. These algorithms have major weaknesses at some respects of Communication and synchronization. The development of network and distributed technology makes cloud computing a reality in the implementation of data mining algorithm. Recently the emergence of cloud computing platform is right able to overcome the storage and calculation bottlenecks of massive data processing, so that huge amount of data mining has become possible. Hadoop is an open source project of Apache for building cloud computing platform. Hadoop framework will help us implementation of clusters easier, faster and more effective[2]. The traditional association rule mining algorithm used in the cloud computing platform is the core problem of massive data mining, the traditional association rule mining algorithm is only suitable for analyzing and mining the centralized data. To make Hadoop applied to traditional association rule mining, a key question is how to parallel the traditional association rule mining algorithms. Then it will complete data mining tasks more efficient in parallel way. But traditional association rule mining algorithms are hard to be paralleled. These association rule mining algorithms will fail if we use the existing centralized data mining methods. Considering above problems, in this paper, we improve Apriori algorithm in order to combine it with the MapReduce programming model of cloud computing and mine the association rules from the mass data in parallel. An improved apriori algorithm: AprioriMR algorithm of association rule based on cloud computing platform was Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Fei Long, Yufeng Zhang, Lv Bin

proposed. The AprioriMR algorithm is well adapted to the Hadoop's MapReduce computing model. It inherits the MapReduce scalability to huge datasets and to thousands of processing nodes. It can handle massive datasets with a large number of nodes on Hadoop platform. The remainder of this paper is organized as follows. Section II describes the cloud computing and Apriori algorithm in details. Section III giving a detailed description of the AprioriMR algorithm. Section IV experimentally verified the validity of the algorithm. We finish in Section V with conclusions and an outlook on future work.

II.

Related Work

II.1.

Apriori Algorithm

Association rules are defined as statements of the form

{X1 , X 2 ,… , X n } → Y , which means that Y present in the

transaction if X1 , X 2 ,… , X n are all in the transaction. There can be a set of items, not just a single item. The probability of finding Y in a transaction with all X1 , X 2 ,… , X n is called confidence. The threshold that a rule holds in all transactions is called support. The level of confidence that a rule must exceed is called interestingness [3]. As one of the most influential algorithms of data mining, the Apriori algorithm discovers frequent itemsets through an iteration method which called layer by layer search. The Apriori algorithm is shown in Fig. 1.

Apriori is the one of the most classical algorithms to extract association rules, and it uses a breadth-first search strategy to generate and test candidate itemsets level-wisely. In spite of the various kinds of rules, the algorithm to discover association rules can generally be broken down into two steps: Find all large frequent itemsets-A large itemset is a set of items that exceeds the minimum support. Generate rules from the large itemsets. The algorithm uses an iterative method called layer search to generate (k + l) itemsets from the k-itemsets. Apriori is a traditional algorithm for association rule and designed to operate on database containing transactions. The main idea of Apriori is producing candidate items, and then scanning the database so as to decide whether they meet the count. The algorithm terminates when no further successful extensions are found. However, the traditional Apriori algorithms encounter many difficulties when deal with massive data. For example, the Apriori algorithm need repeatedly scan the transaction database which may be very large and thus the scale of the database is the important factor to the Apriori performance. the traditional Apriori algorithm spends so much time dealing with particularly large number of candidate sets since candidate (k+1)-itemsets are constructed through the self-join of frequent k-itemsets, the Apriori algorithm may generate vast intermediate itemsets. The Apriori algorithm generates Ck +1 , candidates of frequent itemsets of size k+1, from the frequent itemsets of size k. The number of the Ck +1 may be very vase. To get all frequent sets, it needs to repeatedly scan the database. As Apriori employs an iterative approach, which incurs high I/O overhead for scanning, there are certain limitations regarding in handling of transactions in large database. II.2.

Hadoop

Hadoop is a popular open source implementation of MapReduce, a powerful tool designed for deep analysis and transformation of very large data sets. It enables applications to work with thousands of nodes. Hadoop uses a distributed file system called Hadoop Distributed File System (HDFS) which creates multiple replications of data blocks and distributes them on compute nodes throughout a cluster to enable reliable, and has extremely rapid computations to store data as well as the intermediate results. Hadoop schedules map and reduce tasks to distributed resources, which handles many tough problems, including parallelization, concurrency control, network communication and fault tolerance. Furthermore, it performs several optimizations to decrease overhead involved in the scheduling, network communication and intermediate grouping of results [4]. HDFS is the Hadoop system’s main storage platform. HDFS splits files into a number of data blocks and produces multiple copies. The copied data blocks are stored in a different node for the purposes of improving

Fig. 1. Algorithm Apriori



1113


the reliability and efficiency of the file system when the file is read. The structure of HDFS combines two software programs, Namenode and Datanode. Hadoop is based on Master-Slave architecture. The master server is responsible for managing the namespace and file operations of Datanode software, such as creating, deleting, and renaming files. A file is divided into multiple blocks and is stored on a group of Datanode software programs. Datanode software is responsible meeting users’ needs of HDFS, such as reading or writing data and the implementation of block operations, such as create, delete, and copy from the order of Namenode software[5]. The system is designed in such a way that user data never flows through the Namenode. As a data parallel model, MapReduce is a patented software framework introduced by Google to support distributed computing on large datasets on clusters of computers. Known as a simple parallel processing mode, Map-reduce has many advantages: such as, it is easy to do parallel computation, to distribute data to the processors and to load balance between them, and provides an interface that is independent of the backend technology. The use of the MapReduce library expresses the computation as two functions: Map and Reduce. Map, written by the user, takes an input pair and produces a set of intermediate key/value pairs. The Reduce function, also written by the user, accepts an intermediate key and a set of values for that key. It merges together these values to form a possibly smaller set of values. This allows us to handle lists of values that are too large to fit in memory. Today MapReduce framework is increasingly becoming a popular programming paradigm for data intensive computing, especially when there is ad-hoc data to be processed[6].

III. AprioriMR Algorithm The association rule, based on the degrees of support and confidence for the choice of useful rule, is one of the important means for data mining. The Apriori algorithm may produce a large number of candidate frequent itemsets. For example, if there are 104 frequent 1-itemset, Apriori algorithm needs to produce as many as 107 2-itemsets of candidate. Apriori algorithm may need to be repeated, the Apriori algorithm is only suitable for small data quantity in the single data mining. A single data storage capacity can’t meet the huge web data TB level storage requirements, and with web data increasing, single storage capacity cannot extend, so in the space complexity of single Apriori algorithm can’t meet the web data mining. The association rule mining algorithms will fail if we use the existing centralized data mining methods. But the Hadoop file system is a distributed file system, this model has resulted in the Apriori mining algorithm in Hadoop computing platform can’t be carried on. Because the Apriori algorithm in data processing of integral number according to the set for centralized processing. When the


data stored in the distribution File system, transaction set transaction dispersed in Datanode machines, using Apriori algorithm, each machine Map operation of the iterative calculation, the last calculated in each machine produced frequent item set, and then output to the reduce operation protocol, to obtain the whole data of the frequent itemsets, the frequent item set calculation is obviously not through all transaction set iteration, but in a single Datanode stored in the local storage transaction set iteration, the frequent itemsets is only partial results, not global data results, so Apriori algorithm used in Hadoop produce incorrect results. For the deficiency of traditional Apriori algorithm, to solve these problems, in this paper, we improved Apriori algorithm based on cloud computing platform Hadoop. AprioriMR, an association algorithm based on cloud computing is proposed. Using MapReduce for frequent itemsets counting in apriori algorithm has good scalability. However, repeated scanning of dataset is still needed. The massive data and mining tasks will be decomposed on multiple computers parallel processed. The compute nodes respectively find their local frequent itemsets, then statistics by the Master of the total number of global support, frequent itemsets and ultimately determine the global frequent itemsets, which can significantly increase efficiency of mining algorithm. AprioriMR eliminate the need to iterative scanning of the data to find all frequent items. AprioriMR repeats scanning other intermediate data that usually keep shrinking per iteration. Number of iterations is same as number of iterations in Apriori. The AprioriMR algorithm consists of two steps: the first is generating all frequent itemsets, the second is generating confident association rule from the frequent itemsets. We deploy MapReduce to parallel the first step and have the data stored in the file in order to deal with in the HDFS. The advantage of the improved AprioriMR algorithm is that it can find all the frequent itemsets by scanning the transaction database once, so the time complexity is significantly lowered. We implement the improved AprioriMR algorithm based on MapReduce programming model of cloud computing environment. We use Hadoop components to perform job execution, workflow information storage and use the files replace the database to store datasets. In the files each line can be seen as a transaction, each item is separated with a space or Tab. The datasets in files are split into smaller segments automatically after stored in HDFS and the map function is executed on each of these data segments. At first, the candidate sub item are put out with the counts number after the execution of map function, then the frequent itemsets are found after the execution of map function. When the transaction database is large, and the transaction records of each divided data subset are similar, Intermediate duplicate will account for a large proportion. Undoubtedly that will consume valuable network resources, increase the delay and reduce the I/O performance. Fig. 2 shows the mining process the AprioriMR algorithm.


1114


It can split a large problem space into small pieces and automatically parallelize the execution of small tasks on the smaller space. We use it to reduce the communication overhead without having to take into account the synchronization operations between nodes. The main differences between such algorithms are scale, communication costs; interconnect speed, and data distribution. After all of these, we succeed in parallel association rule mining algorithm and transplant it to Hadoop platform.

IV.

Fig. 2. Algorithm AprioriMR

According to certain principles, the transaction database is horizontally divided into n data subsets which are sent to N nodes. After stored in HDFS, the datasets are split into smaller segments and then transformed to Datanodes. Specify the minimum support and minimum confidence of the association rules. An iteration of the MapReduce produces a frequent itemset. The iteration continues until there aren’t frequent itemsets further. The map function mainly collects the count of every item in candidate itemsets and the reduce function prunes the candidate itemsets which have an infrequent sub pattern. Each frequent item is generated through one execution of map and reduces function. This algorithm’s advantage is that it doesn’t exchange data between data nodes, it only exchanges the counts. In every scan, each map function generates its local candidate items, then the reduce function gets global counts by adding local counts. Master applies to Namenode for the necessary data files, and access the free node list, and return the Metadata to the Master. Master will send metadata to the algorithm stored node, the algorithm storage node send the Apriori algorithm to the original data nodes. Each node scans its data subset to generate the set of candidate itemset. We mainly address the challenge of using the MapReduce model to parallelize Apriori. Hadoop, the framework, can handle many tough problems, including parallelization, concurrency control, network communication and fault tolerance. The algorithm can take full advantage of what Hadoop can provide. It can be easily applied to many commodity machines to deal with mass data without consider the synchronization problem. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Experiments

In order to test the performance of the strategy designed by us, experiment has done on the Hadoop platform. In our experiments, we used Hadoop version 0.19.2, running on a cluster with 25 machines(2 master, 23 slaves). Each machine has duplicate-core processors (running at 2.60GH) and 2GB memory. This experiment is based on massive activity logs and data of all consumers on Taobao.com, to calculate the correlation of products, using one classical apriori algorithm, it also uses AprioriMR algorithm. The experiment compares the classical Apriori with the new AprioriMR algorithm. We use speedup as criterion to measure our algorithm’s superiority. Using Apache’s MapReduce implementation Hadoop, we have compared the time effect and performance of the improved AprioriMR algorithm on the Hadoop platform using different nodes. Fig. 3 shows the time effect of the classical apriori algorithm and the improved AprioriMR algorithm.

Tradition Apriori AprioriMR 2

10 9 8 7 6 5 4 3 2 1 0 1

3

5

7

9 11 13 15 17 19 21 23 25

Fig. 3. Experimental results

The curves reflect the changes of the time complexity of the algorithm when using the Apriori algorithm and using the AprioriMR algorithm proposed in this paper. The improved AprioriMR algorithm based on cloud computing platform for running time is greatly reduced, this is because the improved algorithm without a lot of mining frequent item of iterative calculation, in calculating the space occupied, the improved algorithm has obvious advantage in a cloud computing platform.


1115


The calculation of the number of transactions in the improved Apriori algorithm in cloud computing platform has a certain effect on execution efficiency. Apparently, under the cloud computing environment, the improved algorithm has better performance on mining frequent itemsets from the mass data, and the data set distribution method can improve the efficiency of the improved algorithm. The results show that under the cloud computing environment, the improved algorithm can effectively mine the frequent itemsets from mass data, and the dataset partition method and distribution method can improve the efficiency of the improved algorithm in the heterogeneous cluster environment.

V.

Authors’ information Center for Studies of Information Resources, Wuhan University, Wuhan 430072, Hubei, China. Fei Long was born in Yueyang, China, in 1983. He finished his Master Degree of computer science from the Department of Software, Hunan University, in 2008. His research interests are data mining, e-commerce. Mr. Long currently is a candidate for Doctor Degree of information science from Wuhan University.

Conclusion

In this paper, we proposed the improved AprioriMR algorithm based on cloud computing. The results show that the improved algorithm can effectively mine the frequent itemsets from mass data, and the dataset partition method and distribution method can improve the efficiency of the improved algorithm in the heterogeneous cluster environment.

Acknowledgements This paper is supported by the MOE (Ministry of Education) Project of Key Research Institute of Humanities and Social Sciences at Universities (NO. 08JJD870225) and the National Natural Science Foundation of China under Grant Nos. 71073121.

References [1]

[2]

[3]

[4] [5]

[6]

R. Agrawal and J. C. Shafer, Parallel mining of association rules, IEEE Transactions on Knowledge and Data Eng., vol. 8, n. 6, pp. 962-969, 1996. J. Dean and S. Ghemawat, MapReduce: simplified data processing on large clusters, Communications of the ACM, vol. 51, n. 1, pp. 107 -113, 2008. Ye Yan-bin, A Parallel Apriori Algorithm for Frequent Item sets Mining, Proceedings of the Fourth International Conference on Software Engineering Research Management and Applications (Page: 87-94 Year of Publication: 2006 ISBN:0-7695-2656-X). Apache Software Foundation, “Apache Hadoop,” Jan.2010, URL http://hadoop.apache.org/. K. Shvachko, H. Kuang, S. Radia, and R. Chansler, The Hadoop Distributed File System, 2010 IEEE 26th Symposium on Mass Storage Systems and Technologies MSST(Page: 1-10 Year of Publication: 2010 ISBN: 978-1-4244-7152-2). H. J. Karloff, S. Suri, A model of computation for MapReduce, ACM Special Interest Group on Algorithms and Computation Theory (Page: 938-948 Year of Publication: 2010 ISBN: 978-0-898716-98-6).



1116


Software Quality Assurance Based on Java Modeling Language and Database Shukun Liu1, Xiaohua Yang2, Jifeng Chen3

Abstract – The base and core of software application is software quality. The software quality assurance is a basic method to solve the problem of software. During the process of program running, if all the relations in the program are satisfied, the state of program running is well. The main skills which based on the dynamic program running trace, the theory of relational database and technology of stored procedure are showed. These technologies which can be used to detect the hiding properties among variants, methods and classes and so on are effective, only when they are used on the SQL server 2005 platform. The main hiding properties in the java program are analyzed and the typical methods detecting typical properties are explained in this paper. The result demonstrates that the detecting method for testing the close relationships among variables in java programs is feasible, and these methods can be widely used for analyzing other programs that can be used in software quality assurance. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Java Modeling Language, Relevance Property, Stored Procedure, Software Quality

I.

Introduction

Nowadays, the computer system is developing with a quick speed. With the fast development of computer project, more and more people will take part in the periods of developing, testing and maintaining and so on. So it is very easy to make the codes and documents in the unanimous state which can lead to the sequence that it is very difficult to test and maintain the programs. In order to improve the software quality, the software testing has been a hot issue. There are many ways which are used to test program. But the software quality cannot be improved obviously. Formalizing the properties of JML (Java Modeling Language) is gradually becoming into a main method of improving the software quality. But the difficult work is how to obtain the formal properties of JML. Study shows us that the getting relevance among variants is an efficient way to solve the problem of making contracts which can be widely used in the domains of program testing, evolution, reconstruction and debugging. Today, how to make sure that whether the software is a correct has been a key problem, for the program running efficient and the correctness play an important role in software using. If the relations among variants can be maintained in the running process, the program correctness can be assured in some extent. Usually, which relation should be maintained in the running process maybe listed clearly in the period of program design. For some reason, the relation always changes a lot. In order to make sure the correctness of programs, the relationships should not be changed.


1117

And detecting the relevance of variants is a good method in judging whether the properties have been changed.

II.

Overview of Java Modeling Language

The whole development process can be obviously improved if we descript the anticipated behaviors of methods or classes in the statement form using JML. Adding the JML symbols in java program can help us find whether the relevance of variants have been maintained, describe the function of program code more preciously and debug the errors of program more efficiently. Java Modeling Language is a symbol language often using in the period of detail design. With JML, the class and method in the java program can be seen in a new way. A great many structures describing the behavior are involved. Such as domain model, quantifier, scope of assertion, pretreatment, post processing, inherited conditions and normal behavior standard, and so on. In every class, the relations among class variants and the relationships among method variants must be maintained. Using the java modeling language, the anticipated functions can be described easily, for the processed of achievement are omitted. The relationships among program variants can be observed in this way.

III. Collection of Program Running Trace At any time, the program running trace should be recorded in the process of program analyzing. So the


Shukun Liu, Xiaohua Yang, Jifeng Chen

dynamic program analyzing must be used in the program running trace collection. In order to run the program, firstly, the proper test cases should be chose. Then, the program running trace can be collected and saved in the database in a special form. Once the program running, a running record can be produced. The whole program trace is consisted of many program records. The methods of software quality assurance based on the detection of relationship are gradually become a main way of generating software specification. But the methods of judging whether the relationship among variants are kept can be classified several kinds. In this paper, the relevance relationship means program invariant which is a property that is true at a particular program point, such as might be found in an assert statement, a formal specification or in a representation invariant.

built-in unary function(absolute value, negation, bitwise complement); (4)Invariants over x + y : any invariant from the list of invariants over a single numeric variable, such as x + y = a ( mod b ) ; (5)Invariants over x − y : as for x + y .

Relevance character over three numeric variables: (1)Linear relationship: z = ax + by + c, y = ax + bz + c ; (2)Functions: z = fn ( x, y ) ,for fn ( x, y ) is a built-in binary function(min, max, multiplication, and, or, greatest common divisor; comparison, exponentiation, floating point rounding, division, modulus, left and right shifts).

III.1. List of Invariants The important characters among the variants and attributes of variants can not only be described as relation of numeric variables, but also can be described as relation of nonnumeric variables, relation of the variables of array and the relation of assembling variables. The invariants are as follows, where x, y, and z are variables, and a, b, and c are computed constants. Relevance character over a Single Numeric Variable: (1)The range of single numeric variable: Small value set: for example, x ∈ {a,b,c} . Range limits: for example, x ≥ a,x ≤ b,a ≤ x ≤ b . (2)The relation of non-equal: Non-constant: x ≠ c (c is a constant). The relation of nonfunctional: x ≠ f ( y ) . Relevance character over a Single Nonnumeric Variable: (1)The range of single numeric variable: Small value set: for example x ∈ {a,b,c} ;

These relationships which are described above can be detected in some algorithms which will be showed bellow. These algorithms which can be achieve great success goals are depended on the program trace files based on the relational database.

III.2. Storage Format of Trace File The technology of collecting program running traces including the design methods of relationship pattern for variants in program is based on the relational database. The theory and technology of relational database must be applied in the process. Program can be seen as a description of algorithm in which the transformation is the running of the program. In order to analyze the relations among program variants, the transformation of all program states must be well known. With the program running, lots of trace files recording program running have produced which must be stored. RDBMS is often used to store the trace of program running in order to gain the accurate program invariants. Now an example of trace store pattern in the instance of java program is given. The structure of java program is a hierarchy structure that is in the form of package →class→method→variant. The structure of java program is showed in Fig. 1.

(2) The relation of non-equal

III.3. Storing Relationship of Trace File

Non-constant: for example, x ≠ c . The relation of nonfunctional: x ≠ f ( y ) . Relevance character over two numeric variables: (1)Linear relationship: y = ax + b ; (2)Ordering comparison: x < y,x ≤ y,x > y,x = y,x ≥ y ; (3)Functions: y = fn ( x ) or x = fn ( y ) , for fn is a


The process of analyzing program is analyzing the specific information of the hierarchy structure of program. It is very helpful to analyze the relation of hierarchy of program if recording the program running information according to the relationship of hierarchy. The relations among classes (the definition of the name of class relationship comply with the class name), the relationship among methods (the definition of the name of method relationship comply with the method name) and the relationship among variants (the definition of the International Review on Computers and Software, Vol. 6, N. 6

1118


name of variants relationship comply with the variants name). In java program, there are many kinds of means. The local variants and the global variants are analyzed in the process of program analyzing. The quantity of the class relationship table is equal to the quantity of class. And the quantity of the method relationship table is equal to the quantity of method. The relation of class, relation of method and relation of variants in the methods can be defined as a suite such as < RC, RM, RIGNC, RJGC, RMLNC,RNLC, RPDNC, RQDC >,in which RC denotes the class relation, RM denotes the method relation, RIGNC denotes the global non-collection relation, RJGC denotes the global collection relation, RMLNC denotes the local non-collection relation, RNLC denotes the local collection relation, RPDNC denotes the domain non-collection variants and RQDC denotes the domain collection variants. With the program running, the state information of the methods and variants are stored in the database. According to the definition of relationship above, a four –level data relation can be formed which divided into class level, method level and variant level in the whole analyzing process. The concrete format of trace file is showed in Fig. 2.

IV.

Several Detecting Algorithms of Relevance Among Variants

There many kinds relevance among variants. In this paper only several typical kinds of algorithms are explained bellow.

IV.1. The Detecting Method of Invariants Over any Variable In the trace file, the values of one variant are collected many times. With the program running, maybe the values of the variants are different with each other, but the variant is same. After many times ‘s program running, for one variant if the value is same ,then the variant can be seen as a const. The key part of detecting algorithm is as follow (Fig. 3 - Algorithm 1).

Program

Package1

Package2

………

Packagen

Class1

Class2

………

Classn

Method1

Method2

………

Methodn

Variant1

Variant2

………

Variantn

Fig. 1. The levels of Java program

Fig. 3. Algorithm 1

IV.2. The Detecting Method of Binary Relation The algorithm bellow describes method of detecting the equal relation and unequal relation. The key part of algorithm is as follow (Fig. 4 Algorithm 2).

IV.3. Detecting Method for Ternary Relation The detecting methods of const relation and binary relation are described above. The detecting method of other types of dependence relations such as ternary relation is similar to the method above. The key parts of algorithm of detecting ternary relation are as follow (Fig. 5 - Algorithm 3). Fig. 2. The format of trace file



1119


V.

Conclusion

The research pays attention to the detecting method of variants relevance in java program. Through the typical detecting methods are described above, which can lead a way to other detecting means. For example, the methods to detect relation of non-equals, the range of single numeric variable, the relation of non-equal, relevance character over two numeric variables, and relevance character over three numeric variables are described in this paper. Those methods can be used to improve the software quality. Our future work is to find the detecting methods of multiple relationships. The only way to software quality assurance and software debugging is judging the relevance of variants. The repeated work can be avoided in software analyzing, designing, testing and coding. At the same time the software quality and efficient can be improved.

Acknowledgements This work is supported by scientific research fund of Hunan Provincial Education Department (A Project Supported by Scientific Research Fund of Hunan Provincial Education Department under Grant No.11B073), Hunan Provincial Natural Science Foundation of China (No.10JJ6092) and Scientific Research Fund of Hunan International Economical University (Project name: The research of program design of contract based on Java modeling language, No. 4).

Fig. 4. Algorithm 2

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

Fig. 5. Algorithm 3

[9]


Christoph Csallner, Yannis Smaragdakis, and Tao Xie, DSDCrasher: A hybrid analysis tool for bug finding, ACM Transactions on Software Engineering and Methodology, vol. 17, n. 2, pp. 81-83, 2008. T. Ramananandro. Mondex, An electronic purse: specification and refinement checks with the Alloy model finding method. Formal Aspects of Computing, vol. 20, n.1, pp.21–39, 2008. M. D. Ernst, Dynamically Discovering Likely Program Invariants, Ph.D. dissertation, University of Washington Department of Computer Science and Engineering, (Seattle, Washington), Aug, 2000. G. T. Leavens et al, Preliminary design of JML: A behavioral interface specification language for Java. ACM SIGSOFT Software Engineering Notes, vol. 31, n. 3, pp.1–38, 2009. A. Darvas, P.Muller, Reasoning About Method Calls in Interface Specifications, Journal of Object Technology, vol. 5, n. 5, pp. 59-85, 2009. Alain Giorgetti , Julien Groslambert, Jacques Julliand, and Olga Kouchnarenko, Verification of class liveness properties with Java Modeling Language, IET Software, vol.2, n.6, pp. 500-514, 2008. Michael D. Ernst, Jeff H. Perkins, Philip J. Guo, The Daikon system for dynamic detection of likely invariants, Science of Computer Programming, vol. 69, n. 3 pp. 35-45. 2007. Taweesup Apiwattanapong, Alessandro Orso, and Mary Jean Harrold, JDiff: A differencing technique and tool for object-oriented programs, Automated Software Engineering Journal, vol. 14, n. 3, pp. 3-36, 2007. H. Lehner, P. Muller, Efficient runtime assertion checking of assignable clauses with data groups. In D. Rosenblum and G. Taentzer, editors, Fundamental Approaches to Software


1120


Engineering, vol. 6013 of Lecture Notes in Computer Science, pp.338-352. Springer-Verlag, 2010 [10] L. Burdy, Y. Cheon, D. Cok, M. Ernst, J. R. Kiniry, G. T. Leavens, An overview of JML tools and applications, Journal on Software Tools for Technology Transfer, vol. 7, n. 3, pp. 212-232, 2005.

Authors’ information 1,3

Department of Computer Science and Technology, Hunan International Economics University, Changsha 410205, China.

Dr. Jifeng Chen was born in 1966 in Hunan province, China. He received his Ph. D. degree in Computer Software and Theory from Xi’an Jiaotong University in Airple, 2006. He is Professor of department of computer science and technology. His research interests currently focused on Software Engineering and Software Testing. And he is vice dean of department of computer science and technology of Hunan International Economics University.

2 Department of Computer Science and Technology, University of South China, Hengyang 421001, China.

Shukun Liu received the M.S. degree in computer science and technology from University of South China, 2007. Currently, he is a researcher at department of computer science and technology of Hunan International Economics University, China. His major research interests include database technology, data mining and software engineering. He has published nearly ten papers in related journals. And he is a member of the ACM. Xiaohua Yang received the Ph.D. degree in computer science and technology of Chinese Academy of Sciences, 1999. Currently, he is Professor of department of computer science and technology in University of South China. And he is vice chancellor of University of South china. His major research interests include database technology, data mining and software engineering. He has published nearly fifty papers in related journals and conferences.



1121


Intrusion Detection Based on Improved GA-RBF and PCA Yuesheng Gu, Yanli Zhu, Peixin Qu

Abstract – The intrusion patter identification is a hot topic in this research area. Using Radial Basis Function (RBF) neural networks to provide intelligent intrusion recognition has been received a lot of attentions. However, improper RBF model design may result in a low detection precision. To overcome these problems, a new intrusion detection approach based on RBF classifier and improved genetic algorithm (GA) and principal component analysis (PCA) is proposed in this paper. To alleviate the complexity of the input vector, the PCA has been employed to eliminate redundant features of the original intrusion data. In addition, the improved GA used energy entropy to select individuals to optimize the training procedure of the RBF. Then, the satisfactory RBF model with proper structure parameters was attained. The efficiency of the proposed method was evaluated with the practical data, and the experiment results show that the proposed approach offers a good intrusion detection rate, and performs better than the standard GA-RBF method. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Intrusion Detection, Radial Basis Function, Improved Genetic Algorithm, Principal Component Analysis

I.

Introduction

The security of computer network is one of the most important issues for all the users. Intrusion may lead to absence of the internet service and even cripple the whole system for weeks. Hence, it is very important to detect the intrusion in time to prevent broken-downs. Advanced machine learning algorithm, including evolution algorithm, intelligent artificial neural network (ANN) and support vector machine (SVM) and so on, are all appear in the intrusion detection of the networks [1]-[2]. Among them, ANN [3] is the most extensive used method. However, ANN detection performance is mainly determined by its structural parameters. It is often difficult to determine the ANN parameters without a large number of trials. Although [4] and [5] using GA to tune the ANN structures to improve the network attack detection accuracy, but without considering the GA individual selection adjustment, and just using the KDD dataset to validate their methods, not for practical applications. Therefore, to improve the GA optimization process and to test the real dataset will have important significance for the ANN based intrusion detection [6]-[10]. In order to solve the above problems, this paper proposed a new intrusion detection method. This method has been marked to achieve Radial Basis Function (RBF)’ parameter optimization using improved GA. Moreover, the PCA has been used to the feature selection. It indicates that the feature selection is very essential in the intrusion detection because the original feature space have many useless features to influence the intrusion identification. Eliminate these redundant ones can enhance the intrusion detection rate significantly. Manuscript received and revised October 2011, accepted November 2011

1122

By using the practical dataset for experimental analysis, the analysis results show that the proposed new method can detect the network attack efficiently and the detection rates is higher than the standard GA based method. This paper is organized as follows. In Section 2, the proposed hybrid intelligent method for network intrusion detection based on the combination of PCA, improved GA and RBF is described. The application of the proposed method is presented for network intrusion detection in Section 3. The performance of the feature selection using PCA, as well as the network intrusion detection performance is described. The effectiveness of the proposed method is valued by analyzing the real data. Conclusions are drawn in Section 4.

II.

New Intelligent Model

Due to the interference of inside and external excitations, the network intrusion is a kind of typical non-stationary signal. The different signal components of the network intrusion data exhibit various characteristics, and make a difference between the normal and intrusions. One of the most important procedures in the network intrusion detection is to find out distinguished features to differentiate the intrusion cases from the large amount date sets. However, it is always difficult in choosing an effective feature. Fortunately, the PCA is a powerful tool to select most distinct features from a large amount of characteristics. Therefore, the PCA has been adopted to improve the feature extraction ability of the original data. Meanwhile, RBFNN is an intelligent approach to deal with non-stationary signal. With its strong learning


Yuesheng Gu, Yanli Zhu, Peixin Qu

ability, the RBF is quite suitable for practical intrusion detection. However, its identification efficiency depends on the structure design. This is why the GA optimization is applied to the RBF design. By doing so, satisfactory RBF detection model can be gotten, and consequently the detection rate can be improved. II.1.

where, pt denotes energy probability at lever t. By the annealing selection, we can derive the individual selection probability [14]: P ( zi ) =

Principal Component Analysis (PCA) X = { x1 , x2 , ..., xn }

e

−( E ( zi )+ β f ( zi ) )

∑ i =1 e n

−( E ( zi )+ β f ( zi ) )

(7)

,

where, P ( zi ) denotes the probability of individual z in

xk ∈ R is an m dimension column, n is total sample number. Suppose the linear transform for X is [11]:

new population, E ( zi ) + β f ( zi ) denotes fitness value of

Given

sample

space

m

Fi = aiT X =

= a1i X 1 +a2i X 2 ," + ani X n , i = (1," n )

(1)

II.3.

Then the covariance matrix for F is that:

C = aiT ∑ a j , ( i, j = 1, 2 ," n )

(2)

According to λV = CV , it can calculate the Eigen value λ and eigenvector V for Eq. (2), so: V=

∑ i =1αi xi n

λ = xk ⋅ CV

individual z. After the energy entropy based individual selection procedure, the link between individuals is connected and hence to maintain the diversity of population. Radial Basis Function (RBF)

Radial basis function (RBF) neural network adopts supervised learning and is good at modeling nonlinear data and can be trained in one stage rather than using an iterative process [15]. The RBF network has a feed forward structure consisting of three layers, as illustrated in Fig. 1.

(3) (4)

Arrange Eigen value λ in the descending order, then the high dimension space X can be transformed in a linear space Y: Y =VT X (5) According to the 85% criteria, select the first p components (p < m) in Y as principal components, thus realize the dimension reduction for data X. II.2.

Improved GA

Standard GA [12] involves the following procedures: coding, selection, crossover, and mutation. In the selection, standard GA only searches individuals with adaptive value. The diversity of the population is constrained, which may lead to premature convergence of the GA optimization [13]. To deal with this situation, the energy entropy based selection is employed to increase the diversity of population. As GA selects according to the individuals' energy, it need to calculate each individual z. The individual energy entropy can be expressed as: ⎛ 1 ⎞ E ( t ) = In ⎜ ⎟ ⎝ pt ⎠

(6)


Fig. 1. Structure of RBF neural network

Input layer – there is one neuron in the input layer for each predictor variable. In the case of categorical variables, N-1 neurons are used where N is the number of categories. Hidden layer – this layer has a variable number of neurons (the optimal number is determined by the training process). Each neuron consists of a radial basis function centered on a point with as many dimensions as there are predictor variables. Summation layer – the value coming out of a neuron in the hidden layer is multiplied by a weight associated with the neuron (W1 , W2 , ...,Wn ) and passed to the summation which adds up the weighted values and presents this sum as the output of the network. The following parameters are determined by the training process: (1) The number of neurons in the hidden layer. (2) The coordinates of the center of each hidden-layer RBF function. (3) The radius (spread) of each RBF function in each dimension.


1123

Yuesheng Gu, Yanli Zhu, Peixin Qu

(4) The weights applied to the RBF function outputs as they are passed to the summation layer. II.4.

The Proposed Detection Model

The improved GA was used to optimize the structural parameters of the RBF. The GA begins with random population defined by the problem at hand. Hereby, the neuron number of the RBF’ hidden layers and their weight coefficients. The operation processing is followed by the series of GA activities, such as encoding, fitness evaluation, selection, and genetic operations (reproduction, crossover and mutation). However, before the GA optimization, the PCA is used to reduce the dimension of the RBF inputs. By doing so, the training calculation may be decreased. The inputs of the PCA are the feature space of the KDD 99 training data, F41 × 5000, where 5000 is the sample number and 41 is the feature number of each sample. The PCA has been adopted to fuse the 41 features into several representations in a low dimension. The original feature space F was firstly transformed in a linear space Y, and the cumulative percent of each eigenvector was also calculated. Then, the 85% criteria was adopted to check the first several eigenvectors, and select the first p (p ˆy N −1

-2

10

ˆyk −1 < y j ≤ ˆyk ,⎛⎜

⎝

N 2

⎞ ⎠

< k < N ,k ∈ Z ⎟

yj = 0

-3

10

(21)

ˆyi ≤ y j < ˆyi +1 ,⎛⎜ 1 ≤ i > ta, the Eq.(21) can be simplified to: ⎧ 2 ⎛ ⎫ 2 1⎞ * ⎨ − R + ⎜ R + ⎟ T − 1 ⎬ ta + 2 ⎝ ⎠ ⎩ ⎭

(

ta (1 − ts ) + nts (1 − ta )(1 − ts )

)

(22)

+ 2 R + R ta − R = 0

(

n −1

2

(18)

(

n

=

where T * = T δ is the duration time of a collision measured in time slot units δ (T* in Eq.(1) has the same meaning). This is a binary function extremum question at the constrained condition Eq. (16). We substitute Eq.(16) into Eq.(18). Taking the first-order derivation and imposing it equal to 0, we obtain, after some simplifications, the following Equation: T * − T * ( n + 1) ts + T * n (1 − nR ) ts 2 + =0

)

− 2R + R +

)

T * − (1 − ta )(1 − ts ) T * − 1

n +1

)

2

ta =

δ

)

(21)

Resolve Eq. (22), we obtain:

n

(

(20)

)

(

Ptr ( PAP + PSTA ) = T (1 − Ptr ) + Ptr c

− T * − 1 (1 − ts )

ts 2

⎧ ⎫ ( n + 1) ta + ⎪ ⎪1 − + − t t nR 1 a ( a) ⎪ ⎪ = T* −1 ⎨ ⎬ 2 n ( n + 1) ta ⎪+ ⎪ ⎪ 2 ⎡t + 1 − t nR ⎤ 2 ⎪ a) ⎦ ⎭ ⎩ ⎣a (

Seeing Eq. (17), as Tp, δ and Tc are constants, the total saturation throughput Stotal is maximized when the following quantity is maximized:

=

2

⎧ (1 − ta ) nR ⎫ nta − +⎪ ⎪ ⎪ ta + (1 − ta ) nR ta + (1 − ta ) nR ⎪ T* ⎨ ⎬= 2 ⎪+ n (1 − nR ) ta ⎪ ⎪ ⎡t − 1 − t nR ⎤ 2 ⎪ a) ⎦ ⎩ ⎣a ( ⎭

=

TP

n ( n + 1)

Substitute Eq.(20) into Eq.(19) and use ta to replace ts. We get:

ta (1 − ts ) t (1 − ts ) S AP (16) = = a n −1 S STA nts (1 − ta )(1 − ts ) nts (1 − ta )

(1 − Ptr ) δ + Ptr ( PAPTAP + PSTATSTA ) + PcTc

can be approximated

using:

n

R=

n +1

(19)

We can see that if the R = 1 n , i.e. ta = ts , that means AP and all STAs have the same access probability, Eq. (19) can degenerate to ‘Eq. (27)’ in [12]. Under the


2

( 2R

2

+R

)

2

+

1⎞ ⎡ ⎤ (23) ⎛ +4 R 2 ⎢ − R 2 + ⎜ R + ⎟ T * − 1 ⎥ 2⎠ ⎝ ⎣ ⎦

(

(

)

)

−2 R 2 + ( 2 R + 1) T * − 1

Equation (23) gives us a way to find ta to achieve maximum total saturation throughput in the condition of R ratio of downlink and uplink flows. The method is that we calculate the optimal transmission probability ta through Eq. (23) and then get the Optimal Contention Window size CWopt of AP through Eq. (5). We can see the CWopt only depends on R and T*, so the calculation of our optimal CWopt of AP need not estimate the active number of STAs. In addition the goal of Contention Window adjustment to obtain the optimal value CWopt is not only for system throughput maximized but for asymmetry traffic problems. In our method, when the number of active STAs increase or decrease, we only need to monitor the dynamic change of ratio R of the downlink and uplink flows at AP and send control variable to tune the


1148

Zhengyong Feng, Guangjun Wen, Yuesheng Gu

Contention Window size of all the other active STAs to achieve the ratio R again. This algorithm is presented in the next section.

IV.

Optimal CW Adjustment Method

The main process of our Optimal Contention Window Adjustment method is to monitor the throughput ratio of downlink and uplink flows online at AP and then send the updated CW value to all the other active STAs to hold the ratio R. We assume a time interval T as update period, such as a time interval of beacon frame. During the time T at AP we record the number of successful received ACK frames as NACK which denotes the throughput of downlink flows and the number of successful received DATA frames as NDATA which denotes the throughput of uplink flows. We can also know the data frame payload of AP is PAP and STAs is PSTA. Then, R is calculated as: R=

N ACK PAP T N P = ACK AP N DATA PSTA T N DATA PSTA

(24)

In practical systems, this process is easy to be implemented at AP. Then we compare the measured R with R* which is the goal ratio we want to achieve. If | R − R* |≤ a , here a is a little positive parameter, we do not change the current CW value of STAs. If | R − R* |> a , we use Eq.(25) to compute the new CW value of STAs: ⎛ CWold ⎞ CWnew = round ⎜ ⎟ ⎝ R ⎠

(25)

The flow chart of Optimal Contention Window Adjustment method is shown in Fig. 2.

Fig. 2. Flow chart of optimal CW adjustment method


This algorithm executes every T time period. First, there are two parameters to be initialized R*and T*. R* gets value from [1 n ,1] , here n is the number of active stations. If R* sets to 1 n , that means the downlink flows will occupy only 1 n of total system throughput. If R* sets to 1, that means the downlink flows will occupy half of total system throughput, that is to say, the throughput ratio of downlink and uplink flows is 1:1. The value sets outside this interval is meaningless or unfairness. T* is a parameter calculated from the system settings. Based on R* and T*, we can calculate the CWopt value for AP. Then AP measures the throughput ratio R of downlink and uplink flows online and adjusts the CW value of all the other active STAs to achieve R ≈ R* . Each time the updated CW value can be broadcast to all the active STAs through control frames, such as a beacon frame. In the flow chart, for calculating the optimal Contention Window CWopt of AP, we use some approximation to resolve Eq.(19). In Table I we compare the calculated results of analytical model with results of simulations. We can see the approximation error can be ignored. TABLE I COMPARISON OF CALCULATED AND SIMULATED ta (CWopt) Results of Calculated

Results of Simulated

R

ta

CWopt

R

ta

CWopt

0.2 0.4 0.6 0.8

0.0148 0.0260 0.0351 0.0428

134 76 56 46

0.1999 0.3999 0.5989 0.7997

0.0150 0.0260 0.0357 0.0426

132 76 55 47

1

0.0495

39

0.9986

0.0499

39

V.

Validation and Simulation

First, we validate the analytical results, that is, the Optimal Contention Window CWopt is unrelated to the number of active STAs, we establish the IEEE802.11 infrastructure model in ns2 [13]. The Contention Window backoff scheme of AP is modified. All the other STAs use the original BEB scheme. Set the parameters of IEEE802.11, Data Rate = 2 Mbps, Basic Rate = 1 Mbps, slot-time = 20 us, SIFS = 10 us, PIFS = 30 us, DIFS = 50 us, PLCP preamble = 24 bytes. We simulate the total throughput, downlink and uplink throughput versus Contention Window of AP in three scenarios, that is, n=5, 20 and 60 separately, here n is the number of active STAs. In each scenario we keep the ratio R = 1. That means the throughput ratio between downlink and uplink flows is 1:1. There are n UDP flows transmitted from AP to STAs as the downlink traffic and n UDP flows transmitted from each STA to AP as the uplink traffic. The UDP flows have Rate = 1 Mbps and PacketSize = 1000 bytes. The total traffic input to the system is large enough to saturate the network. The throughput vs. CW value of AP is shown in Fig. 3. From Fig. 3, we can see the optimal value of CW of AP which makes system achieve maximum total saturation International Review on Computers and Software, Vol. 6, N. 6

1149


throughput (also downlink and uplink throughput) is as same as the analytical results from previous section (Seeing Table I, R=1). With the number of active STAs changing from 5, 20 to 60, the optimal CWopt value of AP is unchanged and takes the value of 39. We also can see when CW of AP takes the values between 30 and 50, the total saturation throughput is only a little degraded comparing with the optimal value of 39. So in practice we may relax the precision requirement of CWopt of AP to achieve the maximum saturation throughput. Then, we implement our Optimal Contention Window Adjustment method at ns2 platform. The time periodic T of executing algorithm at AP is 5s and R ratio of downlink and uplink flows is 1 and 0.6 respectively. The total simulation time is 300s, from 0s to 100s, there are 10 active STAs in the WLAN, from 100s to 200s, there are 30 active STAs and from 200s to 300s, there are 60 active STAs. We compare our optimal method with the scheme that just tunes the CW value of AP to maintain the ratio of downlink and uplink flows which is adopted by many current researchers [4]. The simulation results are shown in Figs. 4 and Figs. 5. In the situation of Optimal Contention Window Adjustment method, fixing CWopt value of AP (CWopt=39) and adjusting the CWmin value of STAs to satisfy the requirement of downlink and uplink throughput ratio R*=0.6 and R*=1 respectively, the simulation results are shown in Fig. 4(a) and Fig. 5(a). For three simulation time intervals, there are different active STAs, but the total saturation throughput always achieves the maximum value and is independent on the change of number of STAs.

Figs. 4. Optimal CW vs. Non-optimal CW (R*=0.6)

Figs. 5. Optimal CW vs. Non-optimal CW (R*=1) Fig. 3. Total throughput, downlink throughput and uplink throughput vs. CW of AP (n=5, n=20, n=60)

We also can see the convergence rate of adjustment to approach the ration R* is quick. However, in the situation



1150


of Non-optimal adjustment method, fixing CW value of all active STAs (CWmin=31, CWmax = 1023 which is default value in the standard) and adjusting the CW value of AP to satisfy the ratio R*=0.6 and R*=1 respectively, the simulation results are shown in Fig. 4(b) and Fig. 5(b). We can see the total saturation throughput is degraded comparing with our optimal method besides it continuously degrades further with the increasing of number of active STAs. In addition, the convergence rate of adjustment to approach the ration R* are slower than our optimal method and the throughput ratio of downlink and uplink flows is more fluctuating with the time.

[6]

[7]

[8]

[9]

VI.

Conclusion

[10]

In this paper, we developed an Optimal Contention Window Adjustment method to resolve the asymmetry traffic problem in the WLANs. The method can be easily implemented at AP. AP sets it’s Contention Window as CWopt calculated from our optimization algorithm and periodically adjusts the CW value of STAs to approach the specified R*, ratio of downlink and uplink flows. This method is optimal due to its ability to achieve the maximum total saturation throughput. In addition our method is optimal at the asymmetry access scenario and totally different from the previous proposals. The efficiency of the method has been analyzed by Markov chain mathematic model and validated by simulations. The results show that the method can easily allocate the network bandwidth between downlink and uplink flows according to the required ratio and also achieve the system maximum saturation throughput. It resolves the asymmetry traffic problem and does not degrade the network performance.

This work was supported by Science Foundation of Education Bureau of Sichuan Province, Sichuan, China (No.10ZB019).

References

[2]

[3]

[4]

[5]

[12]

[13]


School of Communication and Information Engineering, University of Electronic Science and Technology of China, Chengdu 611731, Sichuan, China.

2 School of Physics and Electronic Information, China West Normal University, Nanchong 637002, Sichuan, China. 3 Department of Computer Science, Henan Institute of Science and Technology, Xinxiang 453003, Henan, China.

Acknowledgements

[1]

[11]

E. Lopez-Aquilera, J. Casademont, J. Cotrina, and A. Rojas, Performance Enhancement of WLAN IEEE 802.11 for Asymmetric Traffic, Proceedings of IEEE International Symposium on Personal, Indoor and Mobile Radio Communications, (Page: 1463, Year of publication: 2005 ISBN: 9783800729098). N.S.P. Nandiraju, H. Gossain, D. Cavalcanti, K.R. Chowdhury and D.P. Agrawal, Achieving Fairness in Wireless LANs by Enhanced IEEE802.11 DCF, Proceedings of IEEE International Conference on Wireless and Mobile Computing, Networking and Communications, (Page: 132 Year of publication: 2006 ISBN: 1-4244-0494-0). L. Du, Y. Bai and L. Chen, High-Efficient InterActive Pair Scheduling (IAPS) scheme for IEEE 802.11 infrastructure WLAN, Proceedings of IEEE Vehicular Technology Conference, (Page: 2484 Year of publication: 2006 ISBN: 1-4244-0062-7). J. Freitag, N .L. S. da Fonseca and J. F. de Rezende, Tuning of 802.11e Network Parameters, IEEE Communications Letters, vol. 10, n. 8, pp. 611-613, 2006. E. Lopez-Aquilera, M. Heusse, Y. Grunenberger, F. Rousseau, A. Duda and J. Casademont, An Asymmetric Access Point for Solving the Unfairness Problem in WLANs, IEEE Transaction on Mobile Computing, vol.7, n.10, pp.1213-1227, 2008. B. A H. S. Abeysekera, T. Matsuda and T. Takine, Dynamic Contention Window Control Mechanism to Achieve Fairness between Uplink and Downlink Flows in IEEE 802.11 Wireless LANs, IEEE Transaction on Wireless Communications, vol. 7, n. 9, pp. 3517-3525, 2008. G. Bianchi, Performance analysis of the IEEE802.11 distributed coordination function, IEEE Journal Selected Areas in Communications, vol. 18, n. 3, pp. 535-547, 2000. ns2 platform, http://www.isi.edu/nsnam/ns/index.html

IEEE Std. 802.11, Wireless LAN medium access control (MAC) and physical layer (PHY) specifications, (IEEE802.11a/b, 1999) (IEEE802.11g, 2003) (IEEE802.11n, 2009). IEEE Std. 802.11e, Wireless LAN medium access control (MAC) and physical layer (PHY) specifications: medium access control (MAC) Quality of Service enhancements, 2005. X. Wang and S.A. Mujtaba, Performance Enhancement of 802.11 WLAN for Asymmetric Traffic Using an Adaptive MAC Protocol, Proceedings of the IEEE Vehicular Technology Conference, (Page: 753 Year of publication: 2002 ISBN: 0-7803-7467-3). J. Jeong, S. Choi and C. Kim, Achieving Weighted Fairness between Uplink and Downlink in IEEE 802.11 DCF WLANs, Proceedings of International Conference on QoS in Heterogeneous Wired/Wireless Networks, (Page: 10 Year of publication: 2005 ISBN: 0-7695-2423-0 ). S.W. Kim, B.S. Kim and Y. Fang, Downlink and Uplink Resource Allocation in IEEE802.11 Wireless LANs, IEEE Transaction on Vehicular Technology, vol.54, n.1, pp.320-327, 2005.


Zhengyong Feng was born in Sichuan, China, in 1978. He received the B.S. degrees in Physics from China West Normal University, Nanchong, China, in 2001 and the M.E. degrees in Communication Engineering from Institute of Electronics, Chinese Academy of Sciences, Beijing, China, in 2004. Since 2004 to 2007, he was a research associate at China West Normal University. From 2007, he was a PhD candidate at School of Communication and Information Engineering, University of Electronic Science and Technology of China, Chengdu, China. His research interests include wireless networks, QoS of Multimedia streaming over wireless networks and cross layer optimization of wireless networks. Guangjun Wen was born in Sichuan, China, in 1964. He received the B.E. and M.E. degrees in Applied Physics from Chongqing University, Chongqing, China, in 1986 and 1995 respectively and the PhD degrees in Electronic Engineering from School of Physical Electronics, University of Electronic Science and Technology of China, Chengdu, China, in 1998. Since 1998 to 2000, He was a post-doctorate of National Lab of Communications Interference, University of Electronic Science and Technology of China, China. From


1151


2000 to 2001, He was a post-doctorate of Network Technology Research Lab, Electronics and Telecommunications Research Institute (ETRI), Korea. From 2001 to 2002, He was a Research Fellow of Nanyang Technological University (NTU), Singapore. From 2002 to 2004, He was a Senior RF Design Engineer Project Manager of Sumitomo Electric Industries (SEI), Japan. From 2004, He was a professor of Communication and Information Engineering, University of Electronic Science and Technology of China, China. His research interests include wireless networks, wireless location identification, wireless sensor networks. Yuesheng Gu was born in 1973. Now, he is an associate professor of Henan institute of science and technology. His current research area is computer network technology and artificial intelligence.



1152


Sound Analysis for Diagnosis of Children Health Based on MFCCE and GMM Chunying Fang1,2, Haifeng Li1, Wei Zhang1, Bo Yu1

Abstract – Sound diagnosis analysis is now attracting more and more researchers in the world. The purpose of this paper is to study the new method of sound diagnosis to distinguish and characterize abnormal auscultation. firstly, this paper proposes a new feature parameter MFCCE based on MFCC, secondly, finds a appropriate classifier for sound diagnosis through experiments which compares with the classifier GMM and SVM through breath recognition, the result is that GMM is better than SVM. Finally, the further experiments are finished in new recorder data set from hospital to prove the above ideas correctly. The experimental results show that the methods have good performance in Nutrition ineffective cry, cold cough, bronchitis breath, pain cry and non-pain cry recognition. Thus, this paper presents a method for a quantitative assessment of children health through sound analysis. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Sound Diagnosis, Mel-Frequency Cepstral Coefficient Entropy (MFCCE), GMM

I.

Introduction

In recent years, there have been increasing concerns about computer-aided diagnosis, including sound diagnosis [1], sound diagnosis is the most important in tradition Chinese medicine, there are two parts research content in it, one is the main techniques and methods, including vitro jets, aerodynamic, sound chart instrument, spectrum analysis, sound sensors and computer sound collection and analysis system, etc. Another is objective study of the clinical application, for example, someone study the sounds generated by breathing in asthma, or distinguish between breath sound from subjects suffering from various diseases or conditions [2],[3], but mostly rely on doctors subjective judgments. The sound feature parameter and classifier are the most important parts of sound diagnosis. Relevant researches have been done by many scholars who come from different countries. Arnon cohen extracted the linear prediction coefficients and energy envelope features to classify seven types of breath sounds[4], G. Charbonneau extracted the mean frequency and the mean amplitude of the spectrum of the breath sound to diagnosis lung disease[5],the MFCC of the heart sound are extracted and dynamic time warping is used to identify heart sound[6],on the other hand, A. Kandaswamy used the neural networks as classifier [7], P.R. Innocent detected and classification biosounds by using SVM and GMM[8]. But people are focus on single ill, few think about compression sound diagnosis, moreover, the technique and methods are not uniform, author hope that a simple and effective method is found to solve above questions.In this paper, three questions are studied.


1153

The first is to choose sound feather parameters, the second is classifier, and the third is to do further experiments about sound diagnosis based on the above method. Finally, people can identify common diseases based on sound such as cold, bronchitis and so on.

II.

Methodology

This paper is structured as follows: in section II, the novel feature MFCCE and classifier are introduced, while in section III, the bronchitis sound signals used in this paper are acquired and pre-processed. In section IV details, the other features extraction to do further experiment with the MFCCE and GMM, and finally section V concludes the paper. II.1.

MFCC and Entropy

Entropy is a measure of the information based on probability statistical model in information theory by Shannon. Shannon formula is defined as: I ( A ) = −logP ( A )

(1)

where I(A) measure information by the event A occurs, called self-information, P(A) is the probability of A occurs. If a randomized trial has N possible outcomes, or a random message has N possible values, P1, P2, ... , PN, is occur probabilities, then the average information of these events: Hi = −

N

∑ Pi log ( Pi )

(2)

i =1


Chunying Fang, Haifeng Li, Wei Zhang, Bo Yu

H is called entropy. In this paper, entropy H i is inserted in the 13, 26, 39 of 39-dimension MFCC. H i can be computed as follows: Pi = Fi / Hi =

∑

Fi

∑ Pi log Pi

(3) (4)

( 2π )

D 2

∑l

1 2

⋅

l ⎧ 1 ⋅ exp ⎨− ( X − µi ) 2 ⎩

(6) −l ⎫ ∑ i ( X − µi )⎭⎬

where, µi is the d-component vector containing the mean of each feature,

where F is any dimension feature of 39-dimension MFCC. When i is from 1 to 12, then H i is inserted in the 13-dimension MFCC. When i is from 13 to 25, then H i is inserted in the 26-dimension place. When i is from 26 to 39, then H i is inserted in the 39-dimension place. The 39-dimension MFCC extract process is as shown in Fig. 1. Firstly, the FFT(Fast Fourier Transform) is applied to the signal frame by frame , then a module is calculated over these results, and the Mel scale filter bank is a series of 26 triangular band pass filtering. Calculate the logarithm and DCT of the power spectrum of the speech signal. Get 39-dimension MFCC through calculating the first and second order derivative of the 12-dimension MFCC. Finally, MFCCE is got 39-dimension by 3 entropy is inserted into the 13, 26 and 39-dimension.

1

Pi ( X ) =

∑l

is the covariance matrix. Thus, the

Gaussian mixture model can be described by the parameters mean vector, covariance matrix and mixture weights. Therefore, a model can be expressed as a triple as follows:

{

λ = wi , µi , ∑ l

} ,i = 1,2,… ,M

(7)

The GMM method uses the EM algorithm to calculate the models (7)[1][9].

III. Bronchitis Experiments Bronchitis breath sound samples are from a boy who is five months old, normal breath sound is captured after three weeks when he is cure. sampling frequency is 16kHz and 16 bit quantization, frame length is 32ms, frame shift is 16ms. Breath sound is different according to different sample location and instruments, in generally, it is divided into aspiratory and expiratory phase, where the aspiratory time and expiratory time are as one unit cycle(see Fig. 2) from many papers.

Fig. 1. MFCCE extracts processing

II.2.

GMM

The parameters of GMMs can be trained by the K-means clustering and expectation maximization (EM) algorithm. Gaussian mixture model is used to describe the mixture Gaussian distribution, the ideology is: any of the probabilistic density distribution can be approximated by the linear combination of a lot of Gaussian density functions (normal distribution), as follows: P( X |λ) =

M

∑ wi pi ( X )

(5)

i =1

where, M is the order of the mixture model, M is 16 in this system; X is d-component random variable, here, X is the feature vector are obtained by extracting MFCCE parameters of every wave; wi is the first i-Gaussian Pi(X) mixture density function weights, the result is 1; Pi(X) meets the following conditions:


Fig. 2. Respiratory cycle

One unit cycle of normal and bronchitis breath is got, three-dimensional spectrums are apparently different in Fig. 3 and Fig. 4, the pictures show that x-axis represents time which is second, y-axis represents frequency which is Hz, the depth of the color represents the intensity, the breath cycle time is also apparently different, so it is possible that judge diseases from breath sound. The total number of Samples is 40, normal and bronchitis are respectively 20, half the normal breath samples are as training set, others are testing data, the bronchitis samples are the same, 39-MFCCE is extracted and choose the GMM and SVM to recognition, c and g are kernel function in SVM. GMM is better than SVM as shown in Table I. So GMM is more fit on sound diagnosis.


1154


brackets; pain cry is crying sound when he is injection in hospital. In the sounds processing, LPCC and MFCC are two characteristic parameters commonly used [10], A comparative experiment was performed to detect the feature effectively by using four features:12 MFCC and 12 LPCC,36 LPCC, 36 MFCC and 39 MFCCE respectively, four sets of sounds are shown in Table II. the classifier is GMM. The frame length is 256, the GMM number is 16, the sound model are calculated for each case in the training step. The classification accuracy was calculated by taking the number of samples correctly classifier, divided by the total number of testing samples, for example, bronchitis has 10 training samples and 10 testing samples, the classification accuracy was calculated by taking the number of samples correctly in 10 testing samples. After that, the best results were obtained by applying MFCCE. In order to prove the proposed method is correct for the different sound diagnosis, the experiment is designed in Table III, six kinds of sound samples are classified by MFCCE and GMM, There are cry, cough, and breath sound. The cause of the cry is nutrition ineffective, pain cry and non-pain cry. Breath is normal breath and bronchitis breath, cough is cold cough. As the Table III shows sound diagnosis has the high accuracy, this method is effective.

Fig. 3. Three-dimensional spectrum of the normal breath

Fig. 4. Three-dimensional spectrum of the bronchitis breath

IV.

Results

In order to evaluate the proposed methods, a number of experiments were performed with samples from the a sound corpus of children who is two to six years old in First Affiliated Hospital of Heilongjiang University of Chinese Medicine, including cold cough, cry and so on. Samples are divided into train and test set randomly, Training set and testing set number is in following

TABLE I THE BREATH SOUND CLASSIFIER COMPARISON BETWEEN GMM AND SVM Gmm(16)

Svm(c=1.0, g=0.00097)

Normal breath(10/10)

100%

80%

Bronchitis breath(10/10)

100%

90%

TABLE II THE FEATURE COMPARE RESULTS Sound(train/test)

Mfcc(12),lpcc(12)

Lpcc(36)

Mfcc(36)

Mfcce(39)

Bronchitis breath(10/10) Cold cough(5/5) Nutrition ineffective cry(2/2) Pain cry(10/10)

80% 100% 0% 0%

90% 100% 50% 100%

40% 60% 100% 90%

100% 100% 100% 100%

Normal Normal bronchitis Cold Nutrition ineffective Non-Pain cry Pain cry

V.

TABLE III CONFUSION MATRIX FOR SOUND DIAGNOSIS RESULTS Nutrition Bronchitis Cold Non-Pain cry ineffective

Pain cry

10 20 2 4 7 1

Conclusion

In this paper, we have proposed the MFCCE to diagnosis the children health status, as shown in Table III. And the better results are obtained by GMM mixture models. Thus I believed that these technologies have a promising potential in sound diagnosis. This methodology could improve diagnosis results correctly about children. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

20

Results 100% 100% 100% 100% 100% 95.24%

At the same time, In the future, we study sound analysis having in mind three goals. The first goal is to provide a large number of and up-to-data corpus of the available diagnosis sound. The second goal is an effective sound diagnosis system can be construct based on the above theory, the third goal is further improved the classifier results and raise the disease classes by applying a


1155


pre-processing stage or finding better method to process sound.

Acknowledgements Our thanks to supports from the Natural Science Foundation of China (61171186), Key Laboratory Opening Funding of MOE-Microsoft Key Laboratory of Natural Language Processing and Speech (HIT.KLOF.20110xx) and Research Fund for the Doctoral Program of Higher Education No.20050213032. The authors are grateful for the anonymous reviewers who made constructive comments.

References [1]

Chunying Fang, Haifeng Li, Lin Ma, Wenxue Hong, Status and development of human health evaluation based on sound analysis, First International Conference on Cellar Molecular Biology Biophysics and Bioengineering, (Page: 66 Year of Publication: 2010 ISBN:978-1-4244-9158-2). [2] Kenneth Anderson, Yihong Qiu, Arthur R Whittaker, Margaret Lucas, Breath sounds asthma and the mobile phone, The Lancet, vol. 358, n. 9290, pp. 1343-1344, 2001. [3] N. Gavriely and D. W. Cugell, Airflow effects on amplitude and spectral content of normal breath sounds, Journal of Applied Physiology, vol. 80, n. 1, pp. 5-13, 1996. [4] Arnon Cohen and Dorotalandsberg, Analysis and automatic classification of breath sounds, IEEE transactions on biomedical engineering, vol. 31, n. 9, pp. 585-590, 1984 [5] G. Charbonneau, J. L. Racineux, M. Sudraud and H. Tuchais, Digital processing techniques of breath sounds for objective assistance of asthma diagnosis, IEEE Transactions on Biomedical Engineering, vol. BME-31, n. 9, pp. 585-590, 1984. [6] Wenjie Fu, Xinghai Yang, Yutai Wang, Heart Sound diagnosis based on DTW and MFCC, 2010 3rd International Congress on Image and Signal Processing (Page: 2920 Year of Publication: 2010 ISBN: 978-1-4244-6516-3 ). [7] A. Kandaswamy, C. Sathish Kumar, R. P. Ramanathan, S. Jayaraman, N. Malmurugan, Neural classification of lung sounds using wavelet coefficients, Computers in Biology and Medicine, vol. 34, n. 6, pp. 523-537, 2004. [8] Bor Jenq Chua, Xue Jun Li, Huy Dat Tran, Study of automatic biosounds detection and classification using SVM and GMM, 2011 IEEE/NIH Life Science Systems and Applications Workshop (Page: 155 Year of Publication: 2011 ISBN: 978-1-4577-0422-2 ) [9] P. Mayorga, C. Druzgalski, R. L. Morelos, O. H. González, J. Vidales , Acoustics based assessment of respiratory diseases using GMM, the 32nd Annual International Conference of the IEEE EMBS (Page: 6312 Year of Publication: 2010 ISBN: 978-1-4244-4124-2 ). [10] Sunita Chauhana, Ping Wang, Chu Sing Lima, A computer-aided MFCC-based HMM system for automatic auscultation, Computers in Biology and Medicine, vol. 38, n. 2, pp. 221-233, 2008.



School of Computer Science and Technology, Harbin Institute of Technology, Harbin150001, Heilongjiang, China. 2 School of Computer and Information Engineering, Heilongjiang Institute of Science and Technology, Harbin150027, Heilongjiang, China.

Chunying Fang is currently working toward the PhD degree at the Computer Science Department, Harbin Institute of technology (HIT), Harbin, China. She started her teaching career in the College of Computer and information engineering in 2002 in Heilongjiang Institute of Science and Technology and promoted as LECTURER and received her M.E. in information science from Jilin University, China in 2007. Her current research Interests are Speech Recognition, Pattern Recognition & Sound Diagnosis. Haifeng Li Doctoral Supervisor, Director of Speech Processing Lab in School of Computer Science and Technology at HIT and IEEE member. He is the Deputy Dean of Honors School now. He got his Doctor's Degree from Electro-Magnetical Measuring Technique&Instrumentation from HIT in 1997 and Doctor's Degree from Computer, Communication and Electronic Science from University of Paris VI, France in 2002. He started the teaching career in 1994 in HIT, promoted as lecturer in 1995 and professor in 2003. From 1997 to 2002, he is engaged in the post-doctoral research at University of Paris VI, and presided the project of Speech Noise Reduction Research for France Telecom. In August 2004, he became the Assistant Dean of School of Software. His research fields are Audio Information Retrieval & Processing, Artificial Neural Networks. He undertakes many projects of National Natural Science Foundation, Provincial and Ministry Science Foundation and has published over 30 papers in journals and conferences at home and abroad. Wei Zhang received his B.E. in Communication Engineering from Daqing Petroleum Institute, Daqing, Heilongjiang Province, China in 2007 and received his M.E. in Automation from Harbin University of Science and Technology, China in 2009. He is currently working towards the PhD degree in Harbin Institute of Technology. His major fields are Audio Signal Processing, Pattern Recognition& Software development. Bo Yu received his B.E. in Computer Science and Technology from Harbin University of Science and Technology, Harbin, Heilongjiang Province, China in 2004 and received his M.E. in Technology of Computer Application from Harbin University of Science and Technology, China in 2007. He is currently working towards the PhD degree in Harbin Institute of Technology. His major fields are: Speech Recognition, Pattern Recognition& Software development.


1156


Edge Detection and Noise Reduction for Color Image Based on Multi-Scale Feng Xiao1,3, Mingquan Zhou2, Guohua Geng1

Abstract – The traditional method of color image edge detection which converts color image into gray ones and processing the converted gray-scale image edge does not take into account the color information in color images and detected result is not so satisfied. The article proposed a multi-scale edge detection algorithm, which polishing filter for color component output, extending gradient vector of the polished image edges; selecting the thresholds of varied multi-scale image edges according to the improved soft-threshold filtering function and reducing noises and performing the weighted 2-norm fusion of edges of different-scale-image. The results show that SNR value is greater than the traditional algorithm results. The proposed method that takes the color space into consideration significantly improves the detection effect of color-image-edge. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved. Keywords: Edge Detection, Edge Fusion, Multi-Scale, Wavelet Transform, Noise Reduction

I.

Introduction

The edge structures of images are often the most important features for pattern recognition. Color images provide richer and more valuable information for visual perception than that of the gray images. Therefore, the edge detection of color images is becoming a key issue. Meanwhile, the conventional color image edge detection methods mainly convert color image to gray scale images, and go back to gray image edge detection method for color images [1] [2]. In the processes of conversion, the color discontinuity is converted into gray scale discontinuity and loses a lot of valuable information, which makes the follow-up processing can merely base on gray-scale images, or map the results of gray images to the color images. Thus, conventional methods increase the processing complexity of the testing and affect the effectiveness of the results [3]. Studies have shown that [4] [5], vector analysis which is introduced to the color image analysis and processing could achieve more fully use of the image color information.; Zenzo took the multi-dimensional vector sum as the basis of judgment to detect edge[6], Dony and Wesolkowski adopted multi-dimensional vector angular differentiation and detected the edges of different color regions [7]. However, they didn’t take account of the direction of the image gradient vector information in color images. Wavelet transform has the ability to detect local abrupt feature, and it shows innate multi-scale characteristics, and true color image multi-scale edge detection which has been in its ascendant and has attracted many scholars to study in this area [8][9]. However, the detect objects of these methods could only limit to gray scale images.


1157

Meanwhile, there are inevitably a variety of noises in images. In this paper, we proposed a novel method that extracts true-color image edge directly from RGB color space via Multi-scale wavelet, then fuses the extracted image edge in different scales, and denoises the processed result and also enhances the detailed information.

II.

Edge Detection Algorithm

Digital color image is usually stored and expressed in RGB color space. Research results show [10], in RGB space, the sub-optimal and very close to optimal results of edge detection can be achieved. II.1.

Multi-Scale Wavelet Transform

A multi-scale version of this edge detector is implemented by smoothing the surface with a dilated convolution kernel. Suppose two-dimensional image is:

( )

f ( x, y ) ∈ L2 ℜ2

(1)

where the kernel function (namely, the Gaussian function) is:

θ = G ( x, y )

(2)

This is computed with two wavelets that are the partial derivatives of:


Feng Xiao, Mingquan Zhou, Guohua Geng

∂θ ⎧ 1 ⎪⎪ψ ( x, y ) = − ∂x ⎨ ⎪ψ 2 ( x, y ) = − ∂θ ∂y ⎩⎪

(3)

{ }

The scale varies along the dyadic sequence 2 j

j∈Z

,

we thus derive that the wavelet transform components are proportional to the coordinates of the gradient vector of f smoothed by θ 2i :

( (

) )

⎛ W 1 f x, y, 2 j ⎞ ⎜ ⎟= ⎜⎜ 2 j ⎟ W f x, y, 2 ⎟ ⎝ ⎠ ⎛ ∂ ⎞ j ⎜ ∂u f ∗ θ 2 ( x, y ) ⎟ ⎟= = 2j ⎜ 1 ⎜ ∂ ⎟ j f ∗ θ 2 ( x, y ) ⎟ ⎜ ⎝ ∂u2 ⎠ G j = 2 ∇ f ∗ θ 2 j ( x, y )

(

II.3.

Extract the Edge Image by VEG

True color images have at least three components, pixels are vectors. In the system, each color point can be interpreted as a vector extending from the origin to that point in the coordinate system (see Fig. 2). B

)

(

(

From inserting a number of zeroes to the base 3 × 3 template, a 5 × 5 template was generated, and the next scale was formed. When the scale was 2 j , the wavelet multi-scale transform inserted 2 j − 1 zero(es) between the adjacent filter coefficients, which continued to decompose and formed a series of new scales.

)

Blue

(4) Magenta

)

( 0 , 0 , 1)

Cyan

White Gray scale

where * is Convolution, and ∇f is the gradient vector of f . We can see that the two modulus of this gradient vector is proportional to the wavelet transform modulus. The algorithm where the details are captured by smoothing the three color channels (R, G, B) respectively through multi-scale wavelet, utilizing different scale features to witness the detailed local changes in the true color images. The novel method provides more useful information than the ones that processes the true color images directly. II.2.

Green

⎛1 ⎜ 0 1⎜ ⎜2 16 ⎜ ⎜0 ⎜1 ⎝

⎛1 2 1⎞ 1⎜ ⎟ 2 4 2⎟ ⎜ 16 ⎜ ⎟ ⎝1 2 1⎠

R

(1,0,0)

(a) 3 × 3 template

Yellow

Fig. 2. Schematic of the RGB color cube

Let C represent an arbitrary vector in the RGB space: ⎡ cR ⎤ ⎡ R ⎤ C = ⎢⎢ cG ⎥⎥ = ⎢⎢G ⎥⎥ ⎢⎣ cB ⎥⎦ ⎣⎢ B ⎦⎥

(5)

This equation indicates that the components of C are simply the RGB components of a color image at a point. We take into account the fact that the color components are a function of coordinates ( x, y ) by using the notation: ⎡ cR ( x, y ) ⎤ ⎡ R ( x, y ) ⎤ ⎢ ⎥ ⎢ ⎥ C ( x, y ) = ⎢ cG ( x, y ) ⎥ = ⎢G ( x, y ) ⎥ ⎢ c x, y ⎥ ⎢ B x, y ⎥ )⎦ ⎣ ( )⎦ ⎣ B(

0 2 0 1⎞ ⎟ 0 0 0 0⎟ 0 2 0 2⎟ ⎟ 0 0 0 0⎟ 0 2 0 1 ⎟⎠

G

Red

Smoothing Filter Selection

In edge detection, the Gaussian function can effectively filter out the small twists and turns on the edge of the image, this polishing method which combined with the color space, can improve the testing results, so the first derivation of the Gaussian function is chosen as the wavelet Transformation function in Fig. 1.

(0 , 1 , 0 )

Black

(6)

For an image of size M × N , there are MN such vectors. The new color images were formed by component images which were the fruit of polishing each component (R, G, B respectively). Suppose C ( x, y ) is the function of the

(b) 5 × 5 template

Figs. 1. Scale template

In this study, the parameter of the Gaussian function

σ = 0.8, template by 3 × 3 . The wavelet transform was achieved through convolution of the filter, and the multi-scale was realized through interpolation to the filter. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

coordinate ( x, y ) . Let r, g and b be unit vectors along the R , G and B axis of color space (Fig. 2), and define the vectors:


1158


∂R ∂G ∂B ⎧ ⎪⎪u = ∂x r + ∂x g + ∂x b ⎨ ⎪v = ∂R r + ∂G g + ∂B b ∂y ∂y ∂y ⎩⎪

II.4. (7)

Let the quantities g xx , g yy and g xy be defined in terms of the dot product of these vectors, as follows: 2

2

g xx = u ⋅ u = uT u =

∂R ∂G ∂B + + ∂x ∂x ∂x

g yy = v ⋅ v = vT v =

∂R ∂G ∂B + + ∂y ∂y ∂y

2

g xy = u ⋅ v = uT v =

2

2

(8)

(9)

(10)

The maximum rate of change of C is gradient, it is called Vector Extended Gradient. Keep in mind that R , G and B , and consequently the g' x , are functions of x and y . It can be shown [7] that the direction of maximum rate of change of C ( x, y ) is given by the angle: 1 2

⎡

⎤ ⎥ ⎢⎣ g xx − g yy ⎥⎦ 2 g xy

(11)

In this study, in order to enhance the detection of edge effects, the amplitude was doubled, the angle was decreased to half size, and better detection results were obtained. The amplitude is:

(

1 ⎤2

)

⎡ g xx + g yy + ⎥ F (θ ) = 2 ⎢ ⎢ + g − g cos θ + 2 g sin θ ⎥ xx yy xy ⎣ ⎦

(

The above method could detect edges in different scales and get the detailed information and local features at different scale levels. At this time, to get a complete image, we need to integrate those detected edges that differ in scales. Giving consideration to provide both positional information at large scale space and location details at small-scale space, we took use of multi-scale edge weighting method to integrate image edge information at each scale, then the required edge obtained. At scale n , the weighted 2-norm fusion of edges of different-scale image which can be defined as:

2

∂R ∂R ∂G ∂G ∂B ∂B + + ∂x ∂y ∂x ∂y ∂x ∂y

θ = arctan ⎢

The Fusion of Multi-Scale Edge

)

(12)

Along with the process, the gradients of the three components ( R , G , B ) were respectively calculated for each scale C ( x, y ) , and the image edge were formed through synthesizing the corresponding three components with each coordinates ( x, y ) . The algorithm in which the details were captured by smoothing the three color channels ( R , G , B ) respectively through multi-scale wavelet, utilizing different scales features to witness the detailed local changes in the true color images, provides more useful information than the method that processes the true color images directly, and it also provides more information for the signal local processing. The algorithm facilitates the accurate understanding of object perception.


Eˆ image =

n ) ∑ wn ⋅ ( Eimage

2

(13)

n =1

where, wn is the Weighting coefficient of the edges at n different scales; E image are images at different scales, Eˆ image is the edge image obtained after the integration.

In the process of integration, edge images of different scales can be obtained through adjusting the weight value. Greater weights were given to small-scale edge image, better details of local information obtained; while greater weights were given to large-scale edge image, better position, contour information reaped. In experiment, when performing the integration, the scale numbers n took the values of 1,2,3,4,5. The corresponding weighted value read: w1 = 0.4 , w2 = 0.2 , w3 = 0.2 , w4 = 0.1 , w5 = 0.1 .

III. Noise Reduction for Edge Information Before conducting noise reduction we added the Gaussian noise σ = 0.05 to the original image. Color image edge is the region where pixel luminance and chrominance vary in a large range. In edge detection, due to weak edges, noise and other reasons, the visual effects had been impacted greatly. Thus, the results of edge detection require threshold processing. Namely, compare the image edge points in each scale with a fixed threshold, a point less than the threshold value is deleted. Threshold selection is a key issue. Oversized threshold leads to too smooth image, which eliminates some of the weak edge of the image; however, undersized threshold would result in poor smoothness of the image which retains weak edges and the final result— extracted edge accompanied by noise. Donoho [11] proposed a unified threshold calculation method as follows: Thr = σ × 2 log ( N )

(14)

of which: N represents the total number of pixels in the image, σ stands for the standard deviation. According to International Review on Computers and Software, Vol. 6, N. 6

1159


estimation of σ , Donoho proposed a robust median estimation [12]:

σˆ = median ( f ( x, y ) ) / 0.6745

(15)

Selecting the median operations comes down to eliminating the impact on the variance estimation from very rare large amplitude signal. The threshold calculated by Donoho is rough. During edge detection process using the extracted threshold, before which multiplied by a coefficient called the threshold coefficient. Because the distinctiveness of the wavelet transform, the image details were differ in each scale and the pseudo-edge of images were usually retained at small scales. Sub-scale threshold was applied to the algorithm, that is, using different thresholds in large-scale and small-scale respectively and the thresholds were gradually reduced as the scale increases. To achieve the image details selection, we could readjust the value of r , expand the threshold value if the threshold was consistent with the edge, make dynamic adjustment of the threshold range, enhance image edge. The threshold function is defined as:

In Figs. 4, (a) is the original image with noise

σ = 0.01 . (c) (e) (j) (i) are image smoothing of different scales with noise, (b) (d) (f) (h) (j) are Images smooth edge detection results at different scales. In Figs. 5, (a) (b) are edge results of the Sobel and Prewitt operator respectively noise σ = 0.01 . (c) is the edge results of the fused image with no noise,(d) is the edge results of the fused detailed-edge-image through the weighted 2-norm.

(a) original image

(b) the first scale edge

Thr = ⎛ ⎛ ⎛ f ( x, y ) / 0.6745 ⋅⎞ ⎞ ⎞ (16) ⎟⎟⎟ = γ × ⎜ median ⎜ median ⎜ ⎜ ⋅sqrt ( 2 ⋅ log ( N ) ) ⎟ ⎟ ⎟ ⎜ ⎜ ⎝ ⎠⎠⎠ ⎝ ⎝

At different scales the value of r are not the same, at large scales, the edge information is not so clear, larger reading value of r given could achieve better edge details; at small scales, the edge information were clearer, so we gave a smaller value to r . In Different scales, the value of r ranged from 0 to 0.12. To make up for the deficiency of hard and soft threshold, we propose a new threshold function as follows: ⎧ 1 ˆf ( i, j ) = ⎪⎨ γ ⎪⎩ f ( i, j )

f ( i, j ) > Thr f ( i, j ) ≤ Thr

(d)the second scale edge

(e) the third scale smoothing

(f)the third scale edge

(g) the fourth scale smoothing

(h) the fourth scale edge

(i) the fifth scale smoothing

(j) the fifth scale edge

(17)

To compare the image pixel threshold of the obtained edge image at different scales to Thr , provided the image pixel threshold was less than Thr , add the threshold value, details thus are enhanced. γ is called enhanced power, the greater γ is, the greater the threshold value reads, the more prominent the details are.

IV.

(c) the second scale smoothing

Experimental Results and Analysis

In Figs. 3, (a) is the original image. (c) (e) (j) (i) are image smoothing of different scales, (b) (d) (f) (h) (j) are Images smooth edge detection results at different scales.

Figs. 3. Multi-scale true color image edge detection results



1160


(a) noise, original image

(b) noise, the first scale edge

(a) sobel

(b) prewitt

(c) Multi-scale, fusion

(d) noise, Multi-scale, fusion

(c) noise, the second scale smoothing (d) noise, the second scale edge Figs. 5. Edge detection results

(e) noise, the third scale smoothing

The higher the value of SNR, the smaller the difference between the images with noise and without noise, the more noise were filtered out; conversely, the smaller the value of SNR, i.e. the less noise were filtered. Table I shows SNR through Sobel, Prewitt, and the proposed algorithm under different noise background, it is clearly seen that the value of SNR in the traditional algorithm are lower, under the same noise background, the value of SNR in this algorithm is significantly higher than in traditional ones.

(f) noise, the third scale edge

(g) noise, the fourth scale smoothing

TABLE I SNR OF EXTRACTED-IMAGE-EDGES UNDER DIFFERENT NOISE (db) SNR Edge detection algorithm 0.01 0.05 0.10 Sobel 14.19 10.78 10.01 Prewitt 14.32 10.91 10.12 Proposed algorithm 25.95 26.05 25.72

(h) the fourth scale edge

V.

(i) noise, the fifth scale smoothing

(j) noise, the fifth scale edge

Figs. 4. Multi-scale true color image edge detection results with noise

To further validate the efficiency of the algorithm in extracting the edge image detail polluted by noise, we took (18) which showed us the SNR to measure the noise immunity of the algorithm: ⎡ ⎢ ⎢ SNR = 10 lg ⎢ M ⎢ ⎢⎣ i =1

⎤ ⎥ i =1 j =1 ⎥ ⎥ N 2 ⎡⎣ g ( i, j ) − f ( i, j ) ⎤⎦ ⎥ ⎥⎦ j =1 M

N

∑∑ g ( i, j )

2

∑∑

(18)


Conclusion

This paper put forward a novel method where the true color image edge was extracted directly through wavelet multi-scale feature in RGB color model and eliminated noise of multi-scaled images. In practical application, if fully processed by small-scale edge detection, the newly formed edge would leads to dimension disaster while forming the image features. Therefore, we adopted large-scale in providing location information and took small-scale in testing details and then blended the information provided by each scale to implement the object recognition. In addition, according to the improved soft threshold filter function, selecting appropriate threshold of the obtained image edges to perform noise reduction. The algorithm was more flexible in choosing the scales and was able to enhance the detected image quality which then merges into the edges of color images; the algorithm had better robustness, thus more conducive to the image


1161


analysis and processing. Yet, problems on how to be done on the quantitative indicators of the quality for true color image edge detection, how to effectively integrate the scale parameter 2 j of the wavelet transformation and the scale parameter σ of Gaussian function are still to be further examined.

Acknowledgements This work was supported by Principal Fund of Xi'an Technological University (Grant No. XGYXJJ0529, XAGDXJJ1119).


Institute of Visualization Technology, Northwest University, Xi'an 710127, Shaanxi, China. 2 School of Information Science and Technology, Beijing Normal University, Beijing 100875, China. 3 School of Computer Science and Engineering, Xi'an Technological University, Xi'an; 710032, Shaanxi, China.

Feng Xiao received his B.Sc. and M.Sc. degrees in computer science from Xi’an Technological University in 2000 and 2003, respectively. After the M.Sc. degree, he works as a teacher at Xi’an Technological University. Since 2006 he is working toward his Ph.D. degree at Northwest University. His research interests include digital image processing, pattern recognition,

References Cheng Jinyong, Fan Yanbin, Song Jie， Structure analysis and improved method of images edge detection based on wavelets transform, Journal of Qingdao University (Natural Science Edition), vol. 18, n. 3, pp. 71-75, 2005. [2] Wei Bingqian, Yanli, Zhou Jiaxiang, et al., Calculation area of reservoir based on technique of edge detection, Computer Engineering and Applications, vol. 44, n. 26, pp. 217- 220, 2008. [3] Wang Aimin, Zhao Zhongxu, Shen Lansun, Multi-scale color edge detection based on vector order prewitt operators, Journal of Image and Graphics, vol. 4, n. 12, 1999. [4] P. E. Trahanias, A. N. Venetsanopoulos, Color edge detection using vector order statistics, IEEE Trans. Image Processing, vol. 2, n. 4, pp. 259-264． [5] I. J. Scharcansk, A. N. Venetsanopoulos, Edge detection of color images using directional operators. IEEE Trans. Circuits and Systems for Video Technology, vol. 7, n. 2, pp. 397-401, 1997． [6] S. D. Zenzo, A not e on the gradient of a multi-image, Computer Visual Graphic Image Processing, vol. 33, n. 1, pp. 116-125, 1986. [7] R. D. Dony, S. Wesolkowski, Edge detection on color images using RGB vector angles, Proceeding of the 1999 IEEE Canadian Conference on Electrical and Computer E-ngineering Shaw Conference Center (Page: 687-692 Year of Publication: 1999 ISBN: 0-7803-5579-2)． [8] Zhu Qiang, Xu J in, Zhang Duo, et al., Wavelet Modulus Maximum Algorithm in Image Edge Detection, Journal of WUT, vol. 30, n. 6, 2008. [9] Yue Sicong, Zhao Rongchun, Zheng Jiangbin, MERF Based Edge Detection with Adaptive Threshold, Journal of Electronics and Information Technology, vol. 30, n. 4, 2008. [10] S. Wesolkowski, M. E. Jernigan, Robert, Comparison of color image edge detectors in multiple color spaces, Proceedings International Conference on Image Processing (Page: 796-799 Year of Publication: 2000 ISBN: 0-7803-6297-7)． [11] Yan Jingwen, Digital Image Processing. Beijing: National Defence Industry Press, 2007, 2:95-97.． [12] D. L. Donoho, I. M. Johnstone, Adapting to unknown smoothness via wavelet shrinkage, Journal of the American Statistical Assoc, vol. 90, n. 12, pp. 1200-1224, 1995. [1]


imageretrieval. Mingquan Zhou is a professor who is the Dean of Information Science and Technology at Beijing Normal University. His research interests are 3D Vision and cultural heritage protection applications and image retrieval. He has edited three monographs and is the author or coauthor of more than 100 scientific publications presented at several international conferences and workshops, journals and is a member of a council in the CCF. Guohua Geng received her Ph. D. degree in computer science from Northwest University in 2002 .She is a professor who is the Vice-dean of Information Science and Technology at Northwest University. She has been active in various national and international research projects ranging from digital image processing and image retrieval to intelligent information

processing.


1162


Robust Collaborative Tracking in a Multi-Camera Surveillance System Weichu Xiao1, Weihong Chen2, Xiongwei Fei2

Abstract – The distributed architecture of a multi-camera surveillance system is given, a data fusion scheme using path model and estimated target shape changes is described, and a collaborative algorithm in distributed multi-camera surveillance system is proposed in order to track people reliably. Based on the results of multi-camera data fusion, the proposed algorithm allocates cameras to objects according to task priorities, the distance between camera and target and the visibility of targets, which is characterized by the system assigning cameras to the visible target with high priority and the nearest distance to the camera. It is called Priority and Distance Algorithm (PDA). The Quality of Service (QoS) function is introduced to describe the system performance. Experiment results show that the proposed method can coordinate cameras to track people reliably and has good robustness. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Multi-Camera, Collaborative, Tracking, Data Fusion

I.

Introduction

In recent years there are more and more intelligent monitoring system used in commercial, legal and military fields. The visual surveillance for dynamic scenes has become a cutting-edge research field in the computer vision with broad application prospects and potential economic values. Rapid advances in technology have made cheap sensors and especially cameras available. In most cases, it is not possible for a single camera to observe the complete area of interest because the camera field of view (FOV) is finite and objects are occluded in realistic scenarios. Therefore, the goals of surveillance have changed from building surveillance system using only single, powerful camera to building surveillance system using multiple cameras. In multi-camera tracking, many problems arise. On the one hand, we have to solve the problems in single camera tracking such as object segmentation, object occlusion, and so on. On the other hand, we have to deal with the new issues that arise when there are multiple cameras in the system. These issues include: when a person from one camera’s FOV to another camera’s FOV, how to identify this person; how to coordinate cameras to track people reliably. In this paper, we focus on the collaborative tracking and data fusion of multiple cameras in the surveillance system. Some of the technical challenges are to: (1) actively control cameras for tracking multiple moving targets collaboratively; (2) fuse information from multiple cameras into scene-level representations. Thus, a collaboration algorithm is proposed by introducing the weight function.


1163

It distributes cameras to targets depending on the task priority, the distance between target and camera, the visibility of targets, which guarantees that the camera can be distributed to the visible target with high priority and nearest distance to the camera first. In order to track targets reliably and seamlessly under lighting change conditions, a data fusion method using path model and estimated target shape changes is developed, which does not require prior knowledge of the tracking system and complete scene model. Experimental results verify the validity and robust of the algorithm.

II.

Related Background

Currently, there are many researches on multi-camera tracking and surveillance systems for detecting and monitoring the activities of people in or outside the room. The surveillance system can be divided into single-camera surveillance system [1], [2] and multi-camera surveillance system [3], [4], [5]. There are two kinds of structure in the multi-camera surveillance systems: centralized and distributed. With the centralized approach, the surveillance system fuses data on a central server [3]. But the server’s bottleneck on computing and communications will appear if a large number of cameras exist in the system. Therefore, the commonly used approach is the distributed system [4], [5]. Javed et al. designed a multi-camera surveillance system [4]. The detection, tracking and classification of moving targets is performed at the single camera level; at the central server layer, the system uses the single camera tracking results to establish the correspondence between views of the same target in


Weichu Xiao, Weihong Chen, Xiongwei Fei

multiple cameras for collaborative monitoring tasks. The VSAM project of Carnegie Mellon University develops an end-to-end surveillance system that uses multiple pan/tilt/zoom cameras to monitor activities over large area [5]. New targets detected in the scene are classified using neural networks, and then the system uses multiple cameras to track targets through the scene cooperatively.Some researches have been done on the collaborative algorithms in the multi-camera surveillance system [6]-[10]. Nummiaro et al. described a method of selecting automatically best view from different cameras that follow the target using means of color-based particle filtering [6]. Snidaro et al. presented an outdoor multi-camera video surveillance system operating under changing weather conditions [7]. A new confidence measure Appearance Ratio (AR) is defined and the system can select the most appropriate cameras to perform specific tasks by comparing the sensor’s ARs. Lim et al. described the design of the scalable and wide coverage visual surveillance system [8].The system creates a scene model by using a predetermined view of the scene and the maximum weight matching algorithm is used to assign cameras to tasks. Henriksson et al. presented a method for optimizing the use of computational resources in a multi-camera based positioning system [8]. Nguyen et al. propose an algorithm to allocate objects to cameras only using the object-to-camera distance while taking into account occlusion [10].The algorithm attempts to assign objects in the overlapping field of views to the nearest camera. In contrast, the proposed collaborative algorithm for tracking introduces the concept of the task priority on the basis of data fusion results and use parameters to coordinate the monitoring tasks in the multi-camera surveillance system so that the tracking performance of the system is enhanced.

III. The Structure of the Multi-Camera Surveillance System The multi-camera surveillance system with distributed architecture is shown in Fig. 1, which consists of many client sides and one server side. Each client with a camera performs moving target detection, tracking and classification from raw video of single camera at nearly frame-rate and sends its tracking results to the central server.The server is responsible for fusing data from multiple single-camera tracking results and cooperating tracking tasks among multiple cameras to achieve robust tracking.

IV.

Data Fusion

In the multi-camera surveillance system not only is the target information based on the single-camera detection and tracking collected, but also all observations must form a coherent description of the dynamic scene to a certain extent. Therefore, it is necessary to identify the same target in different cameras and integrate the observations into an overall target description. The key is to find the observation of the same target in the system. Here the estimated target appearance changes and path model are used to implement data fusion based on the multiple cameras. IV.1. Data Fusion Algorithm Assume that there are r cameras with overlapping fields and n targets in the surveillance system. The set of targets appeared in the surveillance system is represented by using O = {O1 ,O2 ,...,On } . Each tracked target Oi that has the appearance feature A and path feature P ( s, v, c ) may enter into one camera, move out of another camera or across multiple cameras. Therefore, the set of targets is

{

}

also described as O = Oi ( P ( s,v,c ) , A i ∈ N ) .

The specific algorithm of data fusion is as follows: (1) Detect and track moving targets in the single camera, then take their path characteristics to establish the path model. (2) Estimate the shape changes of the target across multiple cameras, and calculate the distance D between the new target and the tested target existing in the database. (3) Analysis the path model, compare the new target with the tested target from the spatial characteristics, speed, features and curvature features. (4) Determine if target appeared newly is the target existed in the database. If the distance D is less than a non-zero constant or the path features of the new target meet the limitations of the speed and space features, the target appeared newly is labeled by the same tag as the tested target, otherwise it is set the new mark. By using the method of combining the path model with shape information, the system can track the moving targets across multiple cameras reliably. It is not necessary for prior knowledge manual inputted, and is adapt to the environment with brightness changes, and does not require a complete scene model of the system. IV.2. Estimated Target Shape Changes In order to model the shape changes of the target from one camera to another, the color histogram is used to describe the target shape. This is mainly used for learning targets’ color changes when moving among the cameras and the communication is established between a pair of cameras to identify the same target moving among different cameras. For two

Fig. 1. Structure of the multi-camera surveillance system



1164


histograms with m-level p and q, where: ⎧ p = { ˆpu }u =1⋅⋅⋅m ⎪ ⎪m ⎪ ˆpu = 1 ⎪ u =1 ⎨ ⎪q = {qˆ u }u =1⋅⋅⋅m ⎪m ⎪ qˆ = 1 u ⎪ ⎩ u =1

The speed of the tested path is calculated. The speed in the path model is modeled by using the Gaussian distribution and judged if the tested speed is abnormal:

∑

τ=

(v − m )

T

' i

p

( Σ )−1 ( v'i − m p )

(5)

(1) where v'i the speed of the tested path is, m p is the mean,

∑

Σ is the covariance matrix of the path velocity distribution. (3) Curvature features The velocity and acceleration are used to calculate the trajectory’s curvature. The curvature is defined as:

The distance is calculated between them: m

D ( p,q ) = 1 − ∑ ˆpi qˆ i

(2)

u =1

2

k=

B j = fij ( Bi )

(3)

where Bi and B j is the brightness value of the image in

IV.3. Path Model We select three parameters that are the spatial characteristics r, the velocity characteristics s and the curve characteristics c to distinguish the path from the space, speed and curve. Therefore, the path is modeled as P ( s,v,c ) . (1) Spatial characteristics The spatial characteristics of the path include two aspects. First, 90% of the points in the tested path are enveloped inside the path. Second, the Hausdorff distance between the average path and tested path trajectory should be less than the Hausdorff distance between two path surrounding borders [13]. If these conditions are satisfied, we think two paths similar in space, otherwise they are abnormal. (2) Speed characteristics Assume a path Pi ( xi , yi ,ti ) , where xi and yi is the x coordinate and y coordinate of the path Pi at time ti. For a path Pi ( xi , yi ,ti ) , the speed is calculated as:

)

2

3

(6)

where x' and y' is the x and y components of the velocity v' respectively. The mean and variance of the curvature k is measured, which is in line with a Gaussian distribution. The irregular movement is detected by comparing the curvature distribution of the tested path and modeled path using the Mahalanobis distance.

V.

Multi-Camera Collaborative Tracking V.1.

camera Ci and camera C j respectively.

2

⎛ x' t 2 + y' t 2 + 1 ⎞ () () ⎟ ⎜ ⎝ ⎠

Equation (2) is the modified Bhattacharyya coefficient with a measure [11]. By analyzing the value D ( p,q ) the color mismatch of two histograms is obtained. Commonly, light intensity changes may appear in different cameras in the scene. The BTF (Bright Transfer Function) is used to solve this problem [12]. Assume that fij is the BTF from camera Ci to camera C j , then:

(

y'' ( t ) + x'' ( t ) + x' ( t ) y'' ( t ) − x'' ( t ) y' ( t )

Multi-Camera Collaborative Algorithm

Each moving target may be tracked by multiple cameras at the same time in the surveillance system. How to select the most suitable camera to track the moving object is very important when resource constraints. Based on the results of data fusion, this paper introduces a weight function for describing the assignment scheme of the tasks in multiple cameras, in which the task selects the best camera at each time automatically according to the value of the weight. Three variables are used in the weight function: (1) Visibility Vij . This is a value measured between 0 and 1 indicating whether camera Ci can view the geographic location with target O j . (2) Priority Pj . This is the priority of the task T j assigned dynamically by the user according to the

results of data fusion. The greater value of Pj means the higher priority. (3) Distance Dij .This is the distance between camera Ci and target O j in the physical space.

The weight function is defined as: ⎛x −x y −y v'i = ⎜ i +1 i , i +1 i ⎝ ti +1 − ti ti +1 − ti

⎞ ⎟ , i = 0 ,1,...,N − 1 ⎠

(4)


(

Wij = Vij k1 Dij + k2 Pj

)

(7)


1165


where k1 and k2 is the tuning constants of Dij and Pj respectively. The value Wij represents the weight of

camera Ci tracking target O j . The bigger Wij is, the greater the possibility camera Ci tracking target O j is. If Wij is zero, the target O j will not be tracked by

camera Ci . In every moment, the task allocation scheme for cameras as follows: (1) For each task, select a camera with the maximum non-zero weight to perform it. (2) Whenever a higher priority task comes, the task with higher priority is performed. (3) After all tasks are assigned, any camera without assigned tasks performs the basic task of single-camera tracking. This scheme automatically allocates cameras in this way: (1) high-priority tasks have high weights. (2) the target selects the camera with better viewpoints of a particular area. Certainly, if the closest camera is occluded, the second closest camera is selected. If a camera is tracking a target that is moving out of its FOV and entering into another camera FOV, the weight associated with the first camera will decrease, while the weight related to the second camera will increase until the second camera performs the monitoring task of the target automatically. Therefore, the task allocation scheme allows the surveillance system to coordinate cameras automatically for track moving targets seamlessly within the scope of monitoring. V.2.

algorithm, the experiment with real video sequences has been implemented in the distribute surveillance system. Fig. 2 shows the topology of cameras in the system with overlapping FOVs, covering most of the scope of the scene. Fig. 3 shows the trajectories of all people in the system obtained from the recorded scenarios. There are three people in the scene: person P1 goes around the circle; person P2 walks near camera C3; person P3 goes back and forth along a straight line. The brightness of three cameras is inconsistent. The brightest C1 is, the darker C2 is. Fig. 4 shows some tracking results using the method proposed in this paper.

Fig. 2. Topology of C1, C2 and C3

The QoS of the Tracking System

In many cases, the performance is needed to be analyzed by comparing the proposed algorithm with other algorithm. To this end, the introduction of a function here, QoS function is introduced for describing the performance of the tracking system. The QoS function is defined as: Qall = QoS ( C1 ,...,Cr ;O1 ,...,Om ) = =

r

∑ ∑ QoS ( Ci ,O j )

(8) Fig. 3. Trajectories of three persons in the scene

i =1 O j ∈O

(

where Qall is the QoS of the whole system, QoS Ci ,O j

)

is the QoS that camera Ci contributes to target O j , O is the collection of all targets tracked by the cameras. The QoS that camera Ci contributes to target O j is Wij . Thus, the target with high priority close to the camera has high QoS. If the object is occluded or lost, the QoS drops to zero.

VI.


Camera 1

To evaluate the performance of the proposed


Camera 2

Camera 3

Fig. 4. Consistent labeling in the tracking system


1166


To illustrate how the collaborative tracking algorithm described above helps the system track people reliably, we compare PDA with another switching algorithm described in the literature [10]. Since the latter method is based on the distance between the camera and the target and the targets’ occlusion to allocate the camera to the target, it is called DOA (Distance and Occlusion Algorithm). Figs. 5 describe the comparison of PDA and the DOA. Figs. 5(a) shows the QoS assigned to P1, and Figs. 5(b) gives the QoS assigned to all people in the system.

algorithm is characterized by using the target visibility, task priority and the distance between the camera and target to distribute cameras to the tracking tasks. In particular, the concept of task priority makes the system coordinate cameras and track people better. The implementation of the algorithm is based on the results of the data fusion in the distributed system. Experimental results show that the proposed algorithm can coordinate multiple cameras to track targets reliably. Future work will consider the problem of computational load. When many cameras and objects are involved in the system, the processing at each camera should be balanced to avoid computational overloading.

Acknowledgements This work was supported by Hunan Science and Technology Project (No.:2011GK3067).

References (a) A person in the scene

[1]

(b) Three persons in the scene Figs. 5. Comparing the QoS in PDA and DOA

When the system tracks a person, the task tracked with high priority has high QoS . At 7th moment, the target is tracked by camera C3. In the following moments, the target is moving away from C3 and the QoS decreases accordingly. Until at 14th moment, the QoS is reduced to a minimum. From 15th moment, the target is tracked by C1 and close to C1 slowly with higher QoS , see Fig. 5(a). Fig. 5(b) shows the system tracking with three people appeared in the scene at different moments. At 70th moments, P2 with high priority appears in the system and is tracked by C2 firstly . Then, other targets are tracked by camera C2 when the client connecting C2 have the capacity of handling more tasks. As shown in Fig. 5, the PDA has higher QoS than the DOA.

VII.

Conclusion

The multi-camera collaborative tracking algorithm in the distributed system is proposed. It introduces a weight function to select the most suitable camera and defines the QoS to evaluate of the performance of the system. The Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

S. Intille , J. Davis , A. Bobick, Real-time closed-world tracking, Proceedings of the IEEE on Computer Vision and Pattern Recognition (Page: 697 - 703 Year of publication: 1997 ISBN: 0-8186-7822-4). [2] I. Haritaoglu, D. Harwood, L. Davis, W4 : Real time surveillance of people and their activities . IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, n. 8, pp. 809-830, 2000. [3] Y. Bar-Shalom, T. E. Fortmann, Tracking and Data Association. New York : Academic Press , 1988 [4] O. Javed , Z. Rasheed , O. Alatas , M. Shah, Knight：A Real Time Surveillance System for Multiple Overlapping and Non-Overlapping Cameras, Proceedings of the IEEE on Multimedia and Expo., Special Session on Multi-Camera Surveillance Systems (Page: 649-652 Year of publication: 2003 ISBN: 0-7803-7965-9). [5] R. Collins , A. Lipton , H. Fujiyoshi and T. Kanade, Algorithms for cooperative multisensor surveillance, Proceedings of the IEEE, vol. 8, n. 10, 1456-1477, 2001. [6] K. Nummiaro , E. Koller-Meier , T. Svoboda , D. Roth , L. Van Gool, Color-based object tracking in multi-camera environments, Proceedings of the 25th Pattern Recognition Symposium (Page: 591-599 Year of publication: 2003 ISBN: 3-540-40861-4). [7] L. Snidaro , R. Niu , P.K. Varshney , G.L. Foresti, Automatic camera selection and fusion for outdoor surveillance under changing weather conditions, Proceedings of the IEEE on Advanced Video and Signal Based Surveillance (Page: 364 - 369 Year of publication: 2003 ISBN: 0-7695-1971-7). [8] S. N. Lim , L.S. Davis , A. Elgammal, A Scalable image-based multi-camera visual surveillance system, Proceedings of the IEEE on Advanced Video and Signal Based Surveillance (Page: 205 - 212 Year of publication: 2003 ISBN: 0-7695-1971-7). [9] D. Henriksson, T. Olsson, Maximizing the use of computer resources in multi-camera feedback control, Proceedings of the 10th IEEE Real-Time and Embedded Technology and Applications Symposium(RTAS) (Page: 360-367 Year of publication: 2004 ISBN: 0-7695-2148-7). [10] N. T. Nguyen, S. Venkatesh , G. West , H. H. Bui, Multiple camera coordination in a surveillance system, Acta Automatica sinica , vol. 29, n. 3, pp. 408-422. [11] D. Comaniciu ， V. Ramesh ， P. Meer, Real-time tracking of nonrigid objects using mean shift, Proceedings of the IEEE on Computer Vision and Pattern Recognition (CVPR) (Page: 142-149 Year of publication: 2004 ISBN: 0-7695-0662-3). [12] O. Javed, K. Shafique, M. Shah, Appearance Modeling for Tracking in Multiple Non-overlapping Cameras, Proceedings of


1167


[13]

[14] [15]

[16]

the IEEE on Computer Vision and Pattern Recognition (CVPR) (Page: 26-33 Year of publication: 2005 ISBN: 0-7695-2372-2). I. Junejo，O. Javed，M. Shah, Multi feature path modeling for video surveillance, Proceedings of the 7th International Conference on Pattern Recognition (ICPR) (Page: 716-719 Year of publication: 2004 ISBN: 0-7695-2128-2). Z. Q. Hou, C. Z. Han, A Survey of Visual Tracking, Acta Automatica sinica, vol. 32, n. 4, pp. 603-617, 2006. S. S. Huang, L. C. Fu, P. Y. Haiao, Region-level motion-based background modeling and subtraction using MRFs. IEEE Transactions on Image Processing, vol. 16, n. 5, pp. 1446–1456, 2007. B. Y. Liu, L. Yang, J. Z. Huang, et al., Robust and fast collaborative tracking with two stage sparse Optimization, Proceedings of 11th European Conference on Computer Vision, (Page: 142-149 Year of publication: 2004 ISBN: 3-642-15560-X).

Weihong Chen received the Master’s degree in computer science and technology from Hunan University of China. Currently, his major research interests include information security and vision monitoring.

Xiongwei Fei received the Master’s degree in computer science and technology from Hunan University of China. Currently, his major research interests include information security and electronic cash.


School of Communication and Electronic Engineering, Hunan City University, Yiyang 413000, Hunan, China. 2 School of Information Science and Engineering, Hunan City University, Yiyang 413000, Hunan, China.

Weichu Xiao received the B. Eng. degree in Electronic and Information Engineering from Hunan Normal University, Changsha, China, in 1998, He has been working towards the M. Eng. degree in Electronics and Communications Engineering at Hunan University, Changsha, China, since March 2010. From September 2011 to July 2012, He has been working towards the domestic visiting scholars in Circuits and Systems at Xidian University, Xi’an, China. His research interests include super-resolution image and pattern recognition..



1168


Comparison between Three Algorithms for Smooth Support Vector Regression Bin Ren1, Huijie Liu1, Lianglun Cheng2

Abstract – Smooth functions can transform the unsmooth support vector regression into smooth ones, and thus better regression results are generated. It is one of key problems to seek better smooth function in this field for a long time. The BFGS-Armijo and Newton-Armijo algorithms have been used to train smooth support vector regression (SSVR), and the latter has faster speed. Newton-PCG algorithm is just enough method for unconstrained problem which has better speed than Newton in theory. On the numerical experimentation of using BFGS-Armijo, Newton-Armijo and Newton-PCG to train SSVR, this paper gives comparison among the three algorithms, and obtains that Newton-PCG has the best result. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Support Vector Regression, BFGS-Armijo Algorithm, Newton-Armijo Algorithm, Newton-PCG Algorithm

I.

Introduction

II.

Smooth function has been widely studied in numerical modeling [1],[7], which, especially in the interest of the authors, has been successfully applied for classification of and regression model fittings in image processed and pattern recognition[1],[5],[6],[8],[9]. Applying smooth function to regression models means to deal with square unsmoothed issue in ε-insensitive loss function while fitting the regression models [9]. According to the basic concept on how to solve classification problem, Lee et al, used pε2 -function as to smoothly approach the target function, and brings forward the ε-insensitive support vector regression model (ε-SSVR) in 2005[9]. Their results show that the effect of ε-SSVR is better than both LIBSVM [10] and SVM light [11] in both regression property and efficiency. The author took series expansion to generate a new class of polynomial smoothing functions in ε-insensitive support vector regression, it is better than pε2 -function in properties [12]. The BFGS-Armij Newton-Armijo method and Newton-PCG method are three popular methods to solve the smooth models. The authors respectively listed the procedures of these three methods to solve Smooth Support Vector Regression (SSVR), and made a comparative study between the two methods to solve SSVR. The numerical results show that Newton-PCG has the best result.


1169

Smooth Support Vector Recognition

Smooth Support Vector Recognition originally is unconstrained mathematical programming with convex and smooth, and is a method for solving Support Vector Recognition quickly. II.1.

Regression Based Data Fitting

First, we discuss the simplest regression problem in 2-dimensional space: Let’s suppose all values x1 ,x2 , ... ,xm from 1 to m, each xi is corresponding with an observed value yi . The purpose is using the designated data set to generate interdependent function y = f ( x ) .

We usually use this way to solve the problem as below: first, to restrict the function y=f (x) in a simple function class in advance, then searching for f (x) that can meet the following conditions in the function class as much as possible: yi = f ( xi ) , i = 1, 2," ,m (1) For easy deal with, we always use linear regression way, i. e., restricting f ( x ) to be linear function f ( x ) = wx + b . Then search for f (x) which can meet the

equation (1): 2 ( yi − f ( xi ) ) = ( yi − wxi − b )2

(2)

Equation (2) is often used to measure the deviation degree between y = f ( x ) = wx + b and yi = f ( xi ) . The


Bin Ren, Huijie Liu, Lianglun Cheng

smaller value is, the less error is and higher efficient it is. So this process can be translated into the following optimized formula. So that we can define w and b in the function f ( x ) = wx + b : m

∑ ( yi − wxi − b ) w,b

min

2

(3)

i =1

Obviously, the regressive formula and solution above can be extended to a normal situation. First, extending data class (1) to data set S:

S = {( x1 , y1 ) ," ,( xm , ym )} ⊆ R × R n

| x |ε = max {0 ,| x | −ε } , as shown in Fig. 1.

Definition the square of ε-insensitive loss function as | x |ε2 , and the positive function x+ as ( x+ )i = max {0 ,xi } .

Data set S = {( x1 , y1 ) ," ,( xm , ym )} ⊆ R n × R , define

matrix A=[ x1 ,x2, ⋅⋅⋅ xm ], xi is n dimensional vector, each xi is corresponding with an observed value yi , obviously A ∈ R m×n , that is:

S=

{( A , y ) | A ∈ R i

i

i

n

}

, yi ∈ R, for i = 1," ,m

(4) The purpose is using the designated data set S to generate a regression function f ( x ) , let f ( x ) predict y

Secondly, the function class which restrict the function y=f (x) (1) above also can be extended to be a real function set F . Generally, there is not only criterion to measure the deviation of regression function y = f ( x ) from

more accurately according to the new input of x. The standard we use is ε-insensitive loss function:

yi = f ( xi ) . We call equation (3) above as quadratic loss

(7)

| y − f ( x ) |ε = max {0 ,| y − f ( x ) | −ε }

function. Of course, other loss function also can be used. If we name loss function as c ( x, y, f ) . The optimized

| x |ε

formula (3) will become minimization formula with empirical risk: m

min f ∈F

∑ c ( xi , yi , f ( xi ) )

(5)

i =1

−ε

Thus, the interdependent function y = f ( x ) can be obtained, i. e., regression function. When solving the optimized formula (5), the first issue is how to choose the function class set F. For the designated normal training data set S (4), we cannot restrict F to be too small function class, such as linear functions will produce large regression error in a model in nature of nonlinearity. On the other hand, F cannot be too large otherwise the regression function will be meaningless. For example, we will obtain the following equation based on data set S (4) when F is the whole real function set: ⎧ y ,x = xi f ( x) = ⎨ i , i = 1, 2" ,m ⎩ 0,x ≠ xi

0

x

+ε

Fig. 1. ε-insensitive loss function x ε

For linear regression case, f ( x ) = wT x + b , where w ∈ R n is a indeterminate vector, b is a indeterminate constant. ε-insensitive linear regression function was shown in Fig. 2，we select two hyperplanes of the margin in a way and we call the distance between these two hyperplanes is ε zone. Only by there are no training points falling into the margin, we can have loss, and the loss is | y − f ( x) | − ε .

(6) f ( x)

Obviously, the regression function is too illogical. Accordingly the key point is how to choose the function class set F, neither too simple nor too complicated. Furthermore, it becomes difficult to choose the right one for the regression function.

×

f ( x ) = wT x + b + ε

× × × × × × × f ( x) = w x + b × × × × f ( x) = w x + b − ε × × ×

T

T

II.2.

Support Vector Regression

For easy analysis, we define the ε-insensitive loss function of independent variable X as x ε ,


x Fig. 2. ε-insensitive linear regression function


1170


For nonlinear regression case f ( x ) = ωϕ ( x ) + b ,

unconstrained optimized issue formula (9):

where ϕ ( ⋅) : nonlinear function. In theory, we can change it into linear regression ones to settle it which according to kernel technique. Standard regression problem is to solve the following minimum problem [11]: n ⎧ * 1 T ξi + ξi* ⎪min Q ω ,b,ξ ,ξ = 2 ω ω + C i 1 = ⎪ ⎪ s.t. ⎪⎪ ⎨ yi − ω ⋅ ϕ ( xi ) − b ≤ ε + ξi ⎪ * ⎪ω ⋅ ϕ ( xi ) + b − yi ≤ ε + ξi ⎪ * ⎪ξi ≥ 0,ξi ≥ 0 ,i = 1, 2,⋅⋅⋅,n ⎪⎩

(

)

∑(

) (8)

min n+1

( w,b )∈R

ξ = (ξ1 ,ξ 2 ,⋅⋅⋅,ξ n ) , ξ = T

*

(

)

T ξ1* ,ξ 2* ,⋅⋅⋅,ξ n*

deviation during the training. The slack variables ξi and ξi* , correspond to the size of this excess deviation for positive and negative deviations, respectively. The first term, ω T ω is the regularized parameter; thus, it controls the

(

)

function capacity; the second term ∑ ξi + ξi* , is the i =1

empirical error measured by the ε-insensitive loss function. The computation of Standard support vector regression is more complicated, because when solving the Optimization problem, you need to solve quadratic programming, especially when the training sample number is increased. The solution will face curse of dimensionality, in result that we can’t train it. Suykens [14] proposes least squares method-support vector machines (LS-SVM) to make the problem comes down to linear equations, and solving linear equations is easier and faster than the quadratic programming. Standard regression problem is to solve the following problem: ⎧ C n 2 1 T ξi ⎪min Q (ω ,ξ ) = 2 ω ω + 2 i =1 ⎪ ⎨ ⎪ s.t. ⎪ y = ω ⋅ ϕ ( x ) + bξ ,i = 1, 2 ,⋅⋅⋅,n i i ⎩ i

(10)

i =1

Pn2ε ( x,k )

is the square of polynomial

approaching function Pnε ( x,k ) for ε-insensitive loss

is the maximum deviation allowed during the training and C ( > 0 ) represents the associated penalty for excess

n

m

∑| Ai w + b − yi |ε2

Polynomial Smooth Approximation Function

Define , ε ≥0

)

Obviously, the | x |ε2 in formula (4) is not derivative, so this target function is not derivative. With the advancement in networking and multimedia technologies enables the distribution and sharing of multimedia content widely. In the meantime, piracy becomes increasingly rampant as the customers can easily duplicate and redistribute the received multimedia content to a large audience. Insuring the copyrighted multimedia content is appropriately used has become increasingly critical.

II.3.

where:

(

1 T C w w + b2 + 2 2

∑

(9)

In addition, Lee et al adds the parameter 1 b 2 into the 2 objective function to induce strong convexity and to guarantee that the problem has a unique global optimal solution. The regression issue can be expressed by below Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

function x ε , reference [14] have following conclusion (eq. (11)): Pn2ε ( x,k ) = ⎧( x − ε ) 2 , ⎪ 2 ⎪ ⎛ ⎛ 1+ k 2 x − ε 2 ⎞ ⎞ ⎪ ( ) ⎜ ⎜ ⎟ ⎟ + ⎪ ⎜ 1 ⎜ ⎟ ⎟ 2 ⎪ ⎜ ⎜ l ⎟ +⎟ ⎪⎪ ⎞ ⎟ ⎟ ⎜ 2k ⎜ n ( 2l − 3) !! ⎛1 + 1 = ⎨ x ≥ +ε ⎜ ⎜− , ⎜ 2 2⎟ ⎟ ⎟ ∑ k ⎪ ⎜ ⎜⎝ l = 2 ( 2l ) !! ⎝⎜ − k ( x − ε ) ⎠⎟ ⎟⎠ ⎟ ⎪ ⎜ ⎟ ⎪ ⎜ x −ε ⎟ ⎪ ⎜+ ⎟ ⎝ ⎠ 2 ⎪ ⎪ 1 1 1 ⎪⎩ − k + ε < x < k + ε 0 , x ≤ − k + ε

III. Experiment and Discussion Two simulated experiments were selected to demonstrate the analytical results, which were run at Matlab7.0 on a personal computer with an AMD X4 620 processor and 2GB of memory. Based on the first order optimality conditions of unconstrained convex minimization problem, our stopping criterion was satisfied when the 2-norm of gradient of the objective function is less than 10−5 .For an observation vector y and the prediction vector ˆy , the 2-norm relative error of two vectors y and ˆy was defined as follows: || y − ˆy ||2 || y ||2

(12)


1171


This relative error used to measure the accuracy of regression. In order to evaluate how well each method generalized to unseen data, we split the entire data set into two parts, the training set and testing set. The training data was used to generate the regression function that is learning from training data; the testing set, which is not involved in the training procedure, was used to evaluate the prediction ability of the resulting regression function. We also used a stratification scheme in splitting the entire data set to keep the “similarity” between training and testing data sets. That is, we tried to make the training set and the testing set have the similar observation distributions. A smaller testing error indicates better prediction ability. We performed tenfold cross-validation on each data set and reported the average testing error in our numerical results. To generate a highly nonlinear function, a Gaussian kernel was used for all nonlinear numerical tests defined as below:

(

)

Table II and Table III are to use the BFGS-Armijo, Newton-Armijo and the Newton-PCG algorithm, the automatically generated training data set on SSVR results.

Fig. 3. Smooth function approximation comparison chart in the case of k=5 and ε=0.3 TABLE I APPROXIMATION ACCURACY OF SMOOTH FUNCTIONS Smooth Function

Approximation Accuracy

pε2 - function

1.3854 / k 2

p12ε

K Ai , ATj = e

− µ||Ai − A j ||22

, i, j = 1, 2 ,3," ,m

(13)

The parameters µ and C were determined by a tuning procedure. Experiment 1: Approximation Experiment In the case of k=5 ， ε=0.3, the smooth function approximation comparison chart is as Fig. 2, then we can see that， Pn2ε ( x,k ) have higher approximation accuracy than pε2 -function with the same K value. Experiment 2: Numerical Experiment

- function

0.0909/k2

p22ε - function

0.0515/k2

p32ε - function

0.0360/k2

p42ε - function

0.0276/k2

2 5ε

0.0243/k2

p - function

Sample space (M, N) in the M and N, respectively, the number of training samples and the number of sample properties; the number of test samples is m = 1000, algorithm accuracy is ξ = 10-3, to take the initial point of zero vector, smooth factor α (k) = 5, each column of data from top to bottom is the algorithm accuracy rate of training samples(%), test samples of the correct rate(%), the algorithm used The total CPU time (s).

TABLE II RESULT COMPARISON OF THE TRAINING SAMPLE SPACE CHANGE Sample Space (M, N)

Algothrithm

BFGS-Armijo

(200,40)

Newton-Armijo

Newton-PCG

BFGS-Armijo

(1000,40)

Newton-Armijo

Newton-PCG


ε-SSVR

98.12 90.08 1.276 98.12 90.10 0.302 98.12 90.12 0.232 92.03 90.16 3. 85 92.03 90.29 1.126 92.03 91.32 0.862

ε-PSSVR

ε-PSSVR

ε-PSSVR

(n=3)

(n=4)

(n=5)

98.12 90.28 1. 279 98.12 90.16 0.349 98.12 90.12 0. 202 92.03 90.74 3.78 92.03 90.86 1.108 92.03 91.48 0.691

98.12 90.27 1.280 98.12 90.15 0.350 98.12 90.12 0. 205 92.03 90.71 3.87 92.03 90.62 1.120 92.03 91.36 0.735

98.12 90.12 1286 98.12 90.13 0.355 98.12 90.09 0.208 92.03 90.69 3.96 92.03 90.54 1.136 92.03 91.35 0.763


1172


TABLE III RESULT COMPARISON OF THE TRAINING SAMPLE DIMENSION CHANGE Sample Space (M, N)

ε-SSVR

Algothrithm

98.12 90.08 1.276 98.12 90.10 0.302 98.12 90.12 0.232 100.00 80.36 1.256 100.00 81.09 1.016 100.00 82.28 0.230

BFGS-Armijo

(200,40)

Newton-Armijo

Newton-PCG

BFGS-Armijo

(200,300)

Newton-Armijo

Newton-PCG

In Table II and Table III the results are algorithms to meet the accuracy required in the circumstances arise. Analysis and comparison of data can be obtained using the BFGS-Armijo, Newton-Armijo and the Newton-PCG algorithm training SSVR the following conclusions: (1) All results are in training to meet the accuracy required in the circumstances arise, indicating the normal algorithm. Using different algorithms and different smooth function generator is basically the same training accuracy. (2) test accuracy: the whole point of view PSSVC (n = 3) classifier with the best generalization ability; and the training algorithm in terms of Newton-Armijo Newton-PCG and training to get comparable performance, but more than BFGS-Armijo training have higher test accuracy. (3) CPU time: using different smooth function from PSSVR (n=3), PSSVR (n=4) to PSSVR (n=5) consumed CPU time is increasing, indicating a smooth polynomial function of computing time proportional to the number with the smooth order; for different algorithms, BFGS-Armijo, Newton-Armijo and Newton-PCG decreasing consumption of CPU time into the relationship, Newton-PCG algorithm with the best.

IV.

Conclusion

BFGS-Armijo, Newton-Armijo and the Newton-PCG algorithm are used to train and efficient algorithm to solve SSVR, the latter two than the former has the absolute speed advantage, among which the optimal rate of Newton-PCG training.

Acknowledgements This work was supported by the Guangdong Natural Science Foundation Project (No.S2011010002144), The Ministry of Education, Guangdong Province, Production


ε-PSSV

ε-PSSV

(n=3)

R (n=4)

R (n=5)

98.12 90.28 1. 279 98.12 90.16 0.349 98.12 90.12 0. 202 100.00 80.49 1.255 100.00 81.86 1.018 100.00 82.48 0.198

98.12 90.27 1.280 98.12 90.15 0.350 98.12 90.12 0. 205 100.00 80.31 1.258 100.00 81. 62 1.022 100.00 82.33 0.215

98.12 90.12 1286 98.12 90.13 0.355 98.12 90.09 0.208 100.00 80.16 1.263 100.00 81.34 1.031 100.00 82.29 0.227

ε-PSSVR

and Research Projects (No.2010B090400457, NO. 2011B 090400269, NO.2011A091000028), the Natural Science Foundation of Dong Guan University of Technology (No.2010ZQ04).

References [1]

X. Xu, R. Law, T. Wu, Support vector machines with manifold learning and probabilistic space projection for tourist expenditure analysis, International Journal of Computational Intelligence Systems, vol. 2, n. 1, pp. 17-23, 2009. [2] Lei Wang, Ruiqing Zhang, Wei Sheng, Regression forecast and abnormal data detection based on support vector regression, Proceedings of the Chinese Society of Electrical Engineering, vol. 29, no. 8, pp. 92-96, 2009. [3] A. Abdulrahman, A. M. Scott, B.T. Theodore, Real-time prediction of order flowtimes using support vector regression, Computers and Operations Research, vol. 35, n. 11, pp. 3489-3503, 2008. [4] Yubo Yuan, Jie Yan, Chengxian Xu, Polynomial smooth support vector machine (PSSVM), Chinese Journal of Computers, vol. 28, n.1, pp.9-17, 2005. [5] C. astroNeto, M. J. Youngseon, J. Myong, AADT prediction using support vector regression with data-dependent parameters, Expert Systems with Applications, vol.36, n. 2, PART 2, pp. 2979-2986, 2009. [6] C. Andreas, V. M. Arnout, Bouligand derivatives and robustness of support vector machines for regression, Journal of Machine Learning Research, vol. 9, pp.915-936, 2008. [7] O. L. Mangasarian, D R Musicant, Successive overrelaxation for support vector machines. IEEE Transactions on Neural Networks, vol.10, n. 8, pp. 1032-1037, 1999. [8] Y. J. Lee, O. L. Mangasarian, SSVM: A smooth support vector machine for classification, Computational Optimization and Applications, vol.22, n.1, pp. 5-21, 2001. [9] Y. J. Lee, W. F. Hsieh and C. M. Huang, ε-SSVR: A smooth support vector machine for ε-insensitive regression, IEEE Transactions on Knowledge and Data Engineering, vol.17, n. 5, pp.5-22, 2005. [10] C.C. Chang, C.J. Lin, LIBSVM: A library for support vector machines, http://www.csie.ntu.edu.tw/~cjlin/libsvm, 2011. [11] J. Platt. Sequential minimal optimization: A fast algorithm for training support vector machines. Technical Report MSR-TR-98-14, Microsoft Research, MA: MIT Press,1998 [12] Bin Ren, Lianglun Cheng, Polynomial smoothing support vector regression, Journal of Control Theory and Applications. vol. 28, n. 2, pp. 261-265, 2011.


1173


Lianglun Cheng, Prof. and doctoral supervisor of the Faculty of Automation at the Guangdong University of Technology. His research interests include network control & integration, information control and embedded intelligent system.

Authors’ information 1 Faculty of Electro, Dongguan University of Technology, Dongguan 523808, Guangdong, China. 2 Faculty of Automation, Guangdong University of Technology, Guangzhou 510006, Guangdong, China.

Dr. Bin Ren received the Ph.D. degree in control theory and control applications from Guangdong University of Technology China. Currently, he is a researcher at Dongguan University of Technology, China. His major research interests include machine vision and image processing. He has published nearly forty papers in related journals. Huijie Liu were born in Henan,China, in 1979. She received the B.S degree in Electronic and Information Engineering in 2000, and received the M.S degree in Physical Electronics in 2003 from South China University of Technology, Guangzhou, China, Where she is studying for Ph.D. degree. And she is also working in Dongguan University of Technology, Dongguan, China. Her research interests include digital signal processing, chaotic communication and semiconductor lasers dynamics.



1174


PCNN-Histogram Based Multichannel Image Segmentation Algorithm Beiji Zou, Haoyu Zhou, Geli Lv, Guojiang Xin

Abstract – A new multichannel image segmentation method based on pulse-coupled neural network (PCNN) and histogram is presented in this paper. The method segments image by utilizing PCNN’s specific feature that the fire of one neuron can capture firing of its adjacent neurons due to their spatial proximity and intensity similarity. The method judges the weight of each channel in the multichannel image by a histogram-likely way firstly. Then it weightedly combines these channels and active the PCNN. During the iterations of PCNN, the method automatically confirms the best iteration times by comparing the maximum of variance ratio of each iteration. At last, the Shannon entropy rule is used to determine the segmentation results. Experimental results show that the proposed method performs well in both results and efficiency. Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved.

Keywords: Multichannel Image Segmentation, Histogram, Pulse-Coupled Neural Network

I.

Introduction

Multichannel information processing has been considered of vital importance recently with the rapid development of GIS, remote sensing, biomedical imaging, multispectral data management, etc. Retrieval and analysis of object specific features from such a diverse range of channel information are essentially complex tasks primarily due to the complexity of underlying data. Color image segmentation is a classical example of multichannel information processing. The primary challenges faced in the segmentation of color images are the variety and enormity of the color intensity gamut along with the processing of the spectral characteristics of the different color components therein [1]-[12]. Many segmentation methods for multichannel image have been developed in recent years. In [6], the authors proposed a multiresolution image segmentation method based on a graph-theoretic approach. The scheme uses a feature-based, inter-region dissimilarity relation between the adjacent regions of the images under consideration. Finally, the regions are grouped to achieve the desired segmented outputs. The grouping strategy however, is dependent on the chosen inter-region dissimilarity relation. In [7], the authors treated image segmentation as a linear problem instead of the eigenvector approach to a graph-partitioning problem. They achieved segmentation out of spectral partitions with a small isoperimetric constant. The choice of an isoperimetric indicator function obviates the requirements of any coordinate information about the graph. Hence, it results in partitions with optimal cardinalities. In [8], the authors used the mean shift analysis (MS) algorithm in searching for the exact estimation of the color cluster centers in color space.


1175

In [9], the authors developed a robust real-time approach for color image segmentation using the MS segmentation and the normalized cut (Ncut) [10] partitioning methods. The method resorts to the Ncut method to optimize the images clustered by the MS algorithm. These methods however, suffer from the shortcomings in the heuristic choice of a threshold eigenvalue for attaining stable segments. In [11], the authors used the MS clustering method to develop an unsupervised multiscale color image segmentation algorithm. The resultant over- segmented images are then merged based upon a minimum description length criterion. Several statistical mixture models have been proposed to suitably estimate the structural distributions of image data. Examples include the Gaussian mixture [11] and the Dirichlet mixture models [12]. The Gaussian mixture is popular since it is isotropic and can represent data distributions by a mean vector and a covariance matrix. However, the Gaussian mixture fails to discover the true structure of non-Gaussian and asymmetric data distributions. In these situations, the Dirichlet distribution, which is a multivariate generalization of the Beta distribution, can be a good choice for modeling data. In the late 1980s, Eckhorn et al. discovered that the midbrain in an oscillating way created binary images that could extract different features from the visual impression when they had studied the cat visual cortex [1][2]. Based on these binary images the actual image is created in the cat brain. Due to this discovery they developed a neural network, called Eckhorn’s model, to simulate this behavior. In the early 1990s, Rybak et al. also found the similar neural behavior based on the study of the visual cortex of the guinea pig and developed a neural network, called Rybak’s model [3]. Because Eckhorn’s model and Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved

Beiji Zou, Haoyu Zhou, Geli Lv, Guojiang Xin

Rybak’s model provided a simple, effective way for studying synchronous pulse dynamics in networks, they was recognized as being very potential in image processing [4]. The above discoveries have paved the way for the generation of pulse-coupled neural network. Then Johnson et al. carried on a number of modifications and variations to improve its performance as image processing algorithms [4]. This modified neural model is called pulse-coupled neural networks (PCNN). The PCNN is a single layer, two-dimensional, laterally connected network of integrate-and-fire neurons, with a 1:1 correspondence between the image pixels and network neurons. This is a neural network that without any training needed. The output images at different iterations typically represent some segments or edges information of the input image. As a new generation of neural network, the PCNN is good at digital image processing and applied in many fields like image filtering, image enhancement, image fusion, object and edge detection, pattern recognition, etc [5]. The rest of this paper is organized as follows. In Section II, the related algorithms used in our method are presented. In Section III, we propose a new multichannel image segmentation method based on PCNN. In Section VI, the proposed method is tested and compared with some existing methods. Section V is the conclusion.

II.

Supplementaries II.1.

PCNN Model

As mentioned above, the PCNN is two-dimensional, single layered, laterally connected neural network of pulse-coupled neurons, which connect with image pixels one to one. Because each image pixel is associated with a neuron of the PCNN, processing the pixels can be translated into processing the corresponding neurons of the PCNN.The PCNN neuron’s structure is shown in Fig. 1. The neuron consists of an input part, linking part and a pulse generator. The neuron receives the input signals from feeding and linking inputs. Feeding input is the primary input from the neuron’s receptive area. The neuron receptive area consists of the neighboring pixels of corresponding pixel in the input image. Linking input is the secondary input of lateral connections with neighboring neurons. The difference between these inputs is that the feeding connections have a slower characteristic response time constant than the linking connections. The standard PCNN model is described as iteration by equations (1)-(5): Fij [ n ] = e −α F F [ n − 1] + VF

∑ wijklYij [ n − 1] + Iij

(1)

kl

Lij [ n ] = e −α L Lij [ n − 1] + VL

∑ mijklYij [ n − 1]

(2)

kl


(

Uij [ n ] = Fij [ n ] 1 + β Lij [ n ]

)

(3)

Yij [ n ] = step (Uij [ n ] − Eij [ n − 1])

(4)

Eij [ n ] = e−α E Eij [ n − 1] + VE Yij [ n ]

(5)

In these equations, I ij is the input stimulus such as the normalized gray level of image pixels in ( i, j ) position, Fij [ n ] is the feedback input of the neuron in ( i, j ) , and Lij [ n ] is the linking item. U ij [ n ] is the internal activity

of neuron, and Eij [ n ] is the dynamic threshold. Yij [ n ] stands for the pulse output of neuron and it gets either the binary value 0 or 1. The input stimulus (the pixel intensity) is received by the feeding element and the internal activation element combines the feeding element with the linking element. The value of internal activation element is compared with a dynamic threshold that gradually decreases at iteration. The internal activation element accumulates the signals until it surpasses the dynamic threshold and then fires the output element and the dynamic threshold increases simultaneously strongly [47]. The output of the neuron is then iteratively fed back to the element with a delay of one iteration. The inter-connections M and W are the constant synaptic weight matrices for the feeding and the linking inputs, respectively, which dependant on the distance between neurons. Generally, M and W (normally W = M ) refer to the Gaussian weight functions with the distance. β is the linking coefficient. α F , α L and α E are the attenuation time constants of Fij [ n ] , Lij [ n ] and Eij [ n ] , respectively. VF, VL and VE denote the inherent

voltage potential of Fij [ n ] , Lij [ n ] and Eij [ n ] , respectively. Clearly, the standard PCNN has many parameters. A good algorithm using PCNN can make each parameter perform its own functions and further finish the task of data processing very well. Hence, we give a brief explanation about the functions of these parameters in the following. For the feeding channel, α F determines the rate of decay of the feeding channel. Bigger α F causes faster decay of the feeding channel. VF can enlarge or reduce the influence from surrounding neurons. Matrix W refers to the mode of inter-connection among neurons in the feeding receptive field. Generally, the size of W denotes the size of the feeding receptive field. The value of matrix element wijkl determines the synaptic weight strength. In most cases, this channel is simplified via α F =0 and VF =0.


1176


Different from the feeding channel, the link channel usually keep itself as it is. The link channel also has three parameters ( α L , VL and M ) that have the same function to the parameters ( α F , VF and W ), respectively. It is noteworthy that the mode of inter-connection should be designed carefully according to the task of data processing (e.g. image denoising), for it has a great effect on the output of PCNN. Usually, the inter-connection employs the Gaussian weight functions with the distance. The linking coefficient β is an important parameter, because it can vary the weighting of the linking channel in the internal activity. Hence, its value is usually depended on different demands. For example, if much influence from the linking channel is expected, β should be given larger value. All neurons often have the same value of β . It is not absolute. Each neuron can have its own value of β . For the pulse generator, α E indicates the rate of decay of the threshold in the iterative process. Because it directly decides the firing time of neuron, α E is a significant parameter. Smaller α E can make the PCNN work in a meticulous way but it will take much time to finish the processing. On the contrary, larger α E can decrease more running time of PCNN. VE decides the threshold value of fired neuron. If expecting that neuron just fires one time, α E should be set a large value.

Use a threshold at level k to dichotomize the pixels into two classes C1 = [1, 2 ," ,k ] and C2 = [ k + 1,k + 2 ," ,L ] ( C1 is object and C2 is background, or vice versa) .The within-class variance σ W2 and between-class variance

σ B2 can be obtained by equations (7) and (8): ⎧⎪

σ W2 = ⎨ ∑ ( i − µ1 ) ni + 2

⎪⎩i∈C1

⎧⎪

σ B2 = ⎨ ∑ ( µ1 − µT ) ni + 2

⎩⎪i∈C1

∑ ( i − µ2 )

i∈C2

2

⎫⎪ ni ⎬ / N ⎪⎭

∑ ( µ2 − µT )

i∈C2

2

(7)

⎫⎪ ni ⎬ / N (8) ⎭⎪

µ1 , µ2 and µT are the average gray level of C1 , C2 and the image respectively. And the maximum of variance ratio is show in equation (9):

η = max ⎡⎣σ B2 / σ W2 ⎤⎦

(9)

Within-class variance indicates the difference of gray level inside object and background. Between-class variance indicates the difference of gray level between object and background. So when η reaches maximum, it indicates a good segmentation result. It reaches the balance that not only promising greater difference between object and background, but also keeping the minor difference insider object and background at the same time. II.3.

Shannon Entropy

The Shannon entropy for gray image segmentation is shown in equation (10): H ( P ) = − P1 ln P1 − P0 ln P0

(10)

P0 , P1 represent the probability of 0 and 1 exported

from Y ( n ) respectively, which Y ( n ) denotes the binary

Fig. 1. PCNN model

II.2.

The Maximum of Variance Ration

Let the pixels of a given gray image be represented in L gray levels [1, 2 ," ,L ] . The number of pixels at level i is denoted by ni and the total number of pixels by N = n1 + n2 + " + nL . The gray-level histogram is normalized and regarded as a probability distribution in equation (6): pi = ni / N , pi ≥ 0 ,

L

∑ pi = 1

(6)

i =1


output of the segmented image. The Shannon entropy represents the amount of information contained in the image. If the image is segmented resulting in bigger Shannon entropy, the segmented image contains more information from original image and keeps more details, and the segmentation result is better. So the Shannon entropy can be employed to select the best segmentation solution from candidates.

III. Image Segmentation Using PCNN A robust and ideal image segmentation method should obtain meaningful results from any image processing system. Horowitz et al. [4] have provided a good


1177


definition of image segmentation. They define image segmentation as the process of decomposing a given image R , into disjoint nonempty regions or sub-images R1 ,R2 ," ,RN such that: N

(1) U Ri = R ; i =1

(2) Ri is connected for i = 1, 2," ,K ; (3) All pixels belonging to Ri are similar based on some meaningful similarity measure P ; (4) Pixels belonging to adjacent regions Ri and R j are dissimilar based on P . The general approach to segment images using PCNN is to adjust the parameters of the network so that the neurons corresponding to the pixels of a given region pulse together and the neurons corresponding to the pixels of adjacent regions do not pulse together. Assume that the image to be segmented consists of regions and is applied as an input to a PCNN. The network neurons pulse based on their feeding and linking inputs. Note that the feeding input to a neuron is equal to the intensity of its corresponding pixel. Due to the capture phenomenon the neurons associated with each group of spatially connected pixels with similar intensities tend to pulse together. Thus, each contiguous set of synchronously pulsing neurons identifies a segment of the image. A segment identified by the PCNN may be a region, part of a region or union of several regions and sub-regions of the image. Ideally, the goal is to choose the network parameters such that each segment exactly corresponds to a complete region in the image. If such parameters exist then excellent segmentation is possible. Color image is a multi-channel image which is not likely a gray image. For a gray image, we can build a PCNN which neurons are one by one corresponding to each pixel in the image. If we use PCNN for each channel separately, pixels in the same position of the image may be segmented to different region in each channel, because each channel has its own understanding of object region and background region, which is different from other channels. To solve this problem, we translate the image from original R , G , B three channels into a new space which can amplify the difference between object and background. Thus we construct operator m with equation (11): m = R × pR + G × pG + B × pB

⎛

L

⎞

2 = ⎜⎜ ∑ nRi ( i − Ravg ) ⎟⎟ / N R . σ HisR 2

⎝ i =1

(12)

⎠

Ravg is average level channel R . In the same way, we 2 2 calculate σ HisG and σ HisB . Then pR , pG and pB are obtained by equations (13)-(15):

(

)

(13)

(

)

(14)

(

)

(15)

2 2 2 2 pR = σ HisR / σ HisR + σ HisG + σ HisB

2 2 2 2 pG = σ HisG / σ HisR + σ HisG + σ HisB

2 2 2 2 pB = σ HisB / σ HisR + σ HisG + σ HisB

By constructing operator m , we amplify the difference between object and background, and we avoided the problem of processing in each channel separately. Use equation (11) for every pixel in the image, we will get a matrix M = {m1 ,m2 ,...,mn } which has same size as the original multichannel image. To segment the image, we use M to active a PCNN which has the same size of M , iterate over equations (1)-(5). After execution of each iteration cycle, the output for each neuron in the network can be obtained, which can be considered as a segmentation result of the original image. Calculating the maximum of variance ratio of current iteration ηi by equation (9) and compare it with the maximum of variance ηi −1 from last iteration. If ηi > ηi −1 , then continue iteration; Otherwise, if ηi ≤ ηi −1 , then stop iteration, calculate the Shannon entropy of the segmentation results got from the iterations, choose the one which has the maximum Shannon entropy as the final segmentation result. The algorithm flowchart is shown in Fig. 2.

(11)

pR , pG and pB are weights of each channel. They are obtained in the following way: for example, suppose that pixels in channel R contains L levels [1, 2," ,L ] . The

number of pixels at level i in channel R is denoted by nRi and the total number of pixels by N R = nR1 + nR 2 + " + nRL . Then we obtained the variance of histogram in channel R by equation (12):


Fig. 2. Algorithm flowchart


1178


IV.


Our experiments were performed on a 2.4GHz Pentium with 256 Mbytes of RAM, Windows XP and Matlab 6. The PCNN parameters are:

α F = 0.12 , α L = 1.95 , β = 0.10 , α E = 1.10 , VF = 0.45 , VL = 0.20 , VE = 25 , ⎡ 0.5 1 0.5⎤ M = W = ⎢⎢ 1 1 1 ⎥⎥ ⎢⎣ 0.5 1 0.5⎥⎦

algorithm [8]. All these methods perform well on color image segmentation. The experimental results are shown in Figs. 3 and Table I. TABLE I PERFORMANCE EVALUATION OF SEGMENTATION ALGORITHM BASED ON OURS, MULTIRESOLUTION, AND MS Experimental Images Algorithm Lena Airplane Baboon Fruits Runtime(s) 0.5936 0.3197 0.4797 0.8027 Ours Shannon entropy 0.6145 0.4680 0.5441 0.5812 Runtime(s) 0.4432 0.2986 0.7678 0.5273 Multiresolution Shannon entropy 0.5592 0.4685 0.5431 0.5743 Runtime(s) 0.4324 0.2291 0.4344 0.3457 MS Shannon entropy 0.6534 0.4535 0.4688 0.5743

We compared our method with multiresolution image segmentation scheme [6] and mean shift analysis (MS)

Original

Ours

Original

Multiresolution Multiresolution

Multiresolution

MS

MS (b) Image Airplane Segmentaion Results

(a) Image Lena Segmentaion Results

Original

Ours

Ours

Original

MS

Multiresolution

(c) Image Baboon Segmentaion Results

Ours

MS

(d) Image Fruits Segmentaion Results

Figs. 3. Segmentation results of color images



1179


V.

[9]

Conclusion

We proposed a PCNN-Histogram based segmentation algorithm for multichannel images by simplifying the parameters of PCNN and introduced an operator for the histogram for sub-channels. The performance of PCNN is tested with color images. The results show that the proposed method has good performance in both results and execution efficiency. The proper setting of the various parameters of the network, such as linking parameters, interconnection matrices and thresholds depending on the input image, can make PCNN image processing more efficient. In general, the PCNN algorithm should be viewed as a ‘preprocessor’ that needs to be combined with other image processing transforms to be a complete system. Future research will be on the combination of PCNN with other image processing transforms to get better segmentation of object from background.

Acknowledgements This research is supported by National Natural Science Funds of China (No.60970098, No.60803024, No.60903136, and No.61173122), Major Program of National Natural Science Foundation of China (No.90715043), and Specialized Research Fund for the Doctoral Program of Higher Education (No. 20090162110055), Fundamental Research Funds for the Central Universities (No.201021200062), Open Project Program of the State Key Lab of CAD&CG, Zhejiang University (No.A1011).

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

R. Eckhorn, H.J. Reitboeck, M. Arndt and P.W. Dicke, Feature linking via synchronization among distributed assemblies: simulation of results from cat cortex, Neural Computing, vol. 2, n. 3, pp. 293–307, 1990. I.A. Rybak, N.A. Shevtsova, L.N. Podladchikova and A.V. Golovan, A visual cortex domain model and its use for visual information processing, Neural Networks, vol. 4, n. 1, pp. 3–13, 1990. I.A. Rybak, N.A. Shevtsova and V.M. Sandier, The model of a neural network visual preprocessor, Neuro Computing, vol. 4, n. 2, pp. 93–102, 1992. S. He, D. Wang and Y. Nin, Application of Pulse Coupled Neural Network in Image Recognition, Proceedings of the 2010 International Conference on Computing, Control and Industrial Engineering (Page: 415 Year of Publication: 2010 ISBN: 978-0-7695-4026-9). Zhaobin Wang, Yide Ma, Feiyan Cheng and Lizhen Yang, Review of pulse-coupled neural networks, Image and Vision Computing, vol. 28, n. 1, pp. 5-13, 2010. S. Makrogiannis, G. Economou, and S. Fotopoulos, A region dissimilarity relation that combines feature-space and spatial information for color image segmentation, IEEE Transactions on Systems, Man, and Cybernetics Part B, vol. 35, n. 1, pp. 44-53, 2005. L. Grady and E. L. Schwartz, Isoperimetric graph partitioning for image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, n. 3, pp. 469-475, 2006. D. Comaniciu and P. Meer, Mean shift: A robust approach toward feature space analysis, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 24, n. 5, pp. 1-18, 2002.


T. Wenbing, J. Hai, and Z. Yimin, Color Image Segmentation Based on Mean Shift and Normalized Cuts, IEEE Transactions on Systems, Man, and Cybernetics Part B, vol. 37, n. 5, pp. 1382-1389, 2007. [10] J. Shi and J. Malik, Normalized cuts and image segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, n. 8, pp. 888-905, 2000. [11] Q. Luo and T. M. Khoshgoftaar, Unsupervised multiscale color image segmentation based on MDL principle, IEEE Transactions on Image Processing, vol. 15, n. 9, pp. 2755-2761, 2006. [12] S. Medasani and R. Krishnapuram, A comparison of Gaussian and pearson mixture modeling for pattern recognition and computer vision applications, Pattern Recognition Letters, vol. 20, n. 3, pp. 305-313, 1999.

Authors’ information School of Information Science and Engineering, Central South University, Changsha 410083, Hunan, China Dr. Beiji Zou was born in 1961. He is the first author of this paper. Current, he is a professor of Central South University. In 1982, Prof. Zou received the bachelor degree in Zhejiang university, China. In 1984, he received the master degree in Tsinghua University, China and in 2001, he received the Ph. D. degree in Hunan university, China. His major research interest includes digital image processing, computer graphics and software engineering. He has published nearly fifty papers in related journals. Haoyu Zhou (Corresponding author) was born in 1979. Now, he is a Doctor candidate in School of Information Science and Engineering, Central South University, China. In 2002, Mr. Zhou received bachelor degree in Hunan university, China, and in 2005 he received master degree in Hunan university, China. His major research interest is digital image processing. He has published nearly ten papers in related journals. Geli Lv was born in 1973, China. Current, she is a Doctor candidate in School of Information Science and Engineering, Central South University, China. In 1997, Miss Lv received bachelor degree in Hunan university, China, and in 2005 he received the master degree in Hunan university, China.Her major research interest is digital image processing. Guojiang Xin was born in 1979, China. In 2002, Mr. Zhou received bachelor degree in Changsha university of science and technology, China, and in 2006 he received master degree in Hunan university, China. Now, he is a doctor candidate in School of Information Science and Engineering, Central South University, China. His major research interest is digital image processing.


1180

1828-6011(201111)6:6;1-S Copyright © 2011 Praise Worthy Prize S.r.l. - All rights reserved