Neuro-Fuzzy Model Based Classification of Handwritten Hindi Modifiers

International Journal of Application or Innovation in Engineering & Management (IJAIEM) Web Site: www.ijaiem.org Email: [email protected] Volume 3, Issue 6, June 2014

ISSN 2319 - 4847

Neuro-Fuzzy Model Based Classification of Handwritten Hindi Modifiers Gunjan Singh1, Dr. Avinash Pokhriyal2, Prof. Sushma Lehri3 1

Assistant Professor, R.B.S.M.T.C.Agra. Associate Professor, R.B.S.M.T.C., Agra 3 Ex-Director, I.E.T., Dr. B.R. Ambedkar University, Agra 2

Abstract Automatic character recognition is one of the most important and interesting topic of pattern recognition field. A lot of work has been done in the area of machine printed character recognition, but recognition of handwritten characters is comparatively difficult and still a subject of active research. In this paper, we present a neuro-fuzzy system for accurate and adequate classification of handwritten Hindi modifiers or matras. Total 1000 handwritten samples of 10 modifiers are collected from 10 different people. System works in six stages—data collection & scanning, preprocessing, normalization, feature extraction, fuzzy rule set creation and classification. System works on fuzzy information and has a layered architecture—input layer, hidden layers and output layer. Function of first hidden or degree of membership (DOM) layer is to generate the degree of membership for all input-output pattern pairs. Generated degree of membership is then used to train the system using backpropagation algorithm to perform final classification of rules and to determine the membership function of generated output for each class. The membership function is used to determine the classified output pattern corresponding to the input pattern. Feature extraction is done by convoluting a 3X3 mask on pre-processed modifier image. At the end, a comparative study of proposed system with existing systems is also done to evaluate the performance. The proposed system has been implemented in MATLAB 2009 environment.

Keywords: Neuro-fuzzy model, classification, fuzzy if-then rules, membership function, backpropagation learning algorithm.

1. INTRODUCTION Handwritten character recognition is a wide area of pattern recognition field which covers all sorts of character recognition in various application domains such as handwriting recognition, signature recognition, postal address processing etc[1]. This area is becoming more and more important as the amount of data captured & stored (either online or offline) is increasing rapidly with advancements in computer technology, and due to huge amount of data manual identification (such as searching, retrieving, processing and maintaining) of documents’ text is becoming tedious and time consuming process. A lot of work has been done in the area of machine recognition of handwritten characters, still problem cannot be considered as solved. One of the major problems is that handwritten characters are vague, nonuniform in nature and may have some level of fuzziness in their appearance. Artificial neural network and fuzzy set theory has been applied effectively either individually or in combined form to solve automatic character recognition problem since long. An artificial neural network (ANN) is an information processing system in which a large number of processing elements, called neurons, are arranged in layers and each neuron is connected to several other neurons by means of directed communication links. Each link is assigned a weight [2]. A neural network learns through training. Learning or training is done by adjusting weights between layers. Learning can be supervised (learning through teacher), unsupervised (learning without teacher) or reinforcement learning. In supervised learning, network is trained to achieve the target output vector by adjusting weights according to a learning law [3]. In this paper, we used backpropagation learning law to train the proposed network [4]-[8]. Human beings are capable to deal with real world information that is fuzzy, partial and approximate because our experience provides us a way to put together and analyze these pieces of information into a structured way. Fuzzy logic is an effective tool to describe human knowledge involving vague or fuzzy uncertainties. The concept of fuzzy set theory was given by Lofti A. Zadeh in 1965. According to Zadeh [9], fuzzy logic is a method to solve and model the uncertainties associated with human thinking and reasoning processes. The concept of fuzzy set allows partial membership so a fuzzy set may have smooth boundaries. Membership of an element of fuzzy set is expressed by the degree of compatibility of the element with the definition of that set. Membership is expressed between number 0 (complete non-membership) and 1(complete membership). Value in-between 0 and 1 represents partial membership [10]. A fuzzy set A in the universe of discourse U (into the interval [0,1]) can be defined as a set of ordered pairs and is given by: Ã = { (x, µ Ã (xi) ) | x ∈U} where µ Ã (x) represents degree of membership of x in Ã and called membership function of Ã. Fuzzy rules is by far the most used and successful method of fuzzy logic theory to represent inexact and ambiguous data.

Volume 3, Issue 6, June 2014

Page 311


ISSN 2319 - 4847

Strength of neural network is its learning and storing capability, while fuzzy logic handles the complexity of fuzzy data very well by expressing it in the form of IF-THEN rules [11]. So, a neural-fuzzy system incorporates the features of learning through logic and hence enhances efficiency of the system. In this paper, we presented a novel neural-fuzzy model to solve classification problem of handwritten Hindi curve script modifiers. In neuro-fuzzy aproach, fuzzy and neural networks are used independently, where each one perform its functions, and serve the system by incorporating and complementing each other in order to accomplish the desired task. System works by interpreting the fuzzy rules in terms of the neural network. Fuzzy sets are taken as weights, while fuzzy rules, input and output variables are taken as neurons[12]. System learns by applying a learning algorithm as in a neural network, and accordingly changes in weights taken place. As a result, change in the architecture may be possible. So networks learns in the form of fuzzy rules. It can be said that, in neuro-fuzzy approach, neural network is used to—(i) process fuzzy inforamation, and (ii) implement the fuzzy system. This paper is organized in six sections. In section 1, a brief introduction of automatic character recognition, artificial neural networks, fuzzy set theory and neuro-fuzzy system is given. Section 2 throws some light on features of Hindi language. Section 3 presents the proposed recognition system. In Section 4 experimental results are presented. Section 5 and 6 are devoted to comparative study & conclusion respectively.

2. HINDI LANGUAGE Hindi is one of the official languages of India. It is an Indo-Aryan language. It is the world’s third most commonly used language after Chinese and English and has approximately 500 million speakers all over the world. It is based on Devnagari script. The basic character set has 39 characters : 13 vowels (SWARS) and 33 consonants (VYANJANS). Language also has 10 basic modifiers (figure 1) . Characters may also have half form, composite form and compound form (figure 2).

(a) (b) Figure 1 (a) Basic character set of Hindi language and (b) set of modifiers

(a) (b) (c) Figure 2 (a) Half characters, (b) composite characters, and (c) compound characters Now-a-days Hindi language is used worldwide in many application areas such as banking, engineering, medical etc. Due to its increasing popularity, automatic Hindi language recognition systems have now become the need of time. Research in this area started in early 1970s. Till than a number of methods and techniques have been proposed [13]-[19].

3. PROPOSED SYSTEM Proposed system works in 6 stages—data collection & scanning, preprocessing, normalization, feature extraction, fuzzy rule set creation and classification. Flow process is shown below--


Page 312


Collect handwritten data of 10 modifiers on paper

ISSN 2319 - 4847

Recognized modifier

Defuzzify output data Scan data to convert data to gray scale images

Create and train fuzzy neural network

Apply median filtering to noise reduction

Create a set of fuzzy rules for each modifier by using extracted feature values

Apply Global thresholding to perform binarization

Normalized modifier image Normalize images to the size of 7x7

Extract values of projection profile, normalized chain code & pixel ratio features for each modifier image

Figure 3 Flow process of proposed system

3.1 DATA COLLECTION & SCANNING Handwritten modifier data samples are collected on paper and scanned through an optical scanner. 3.2 PRE-PROCESSING During pre-processing following operations are applied on data to make it noise free and suitable for further processing – a. Median filtering[20] to reduce noise and false points, b. Global thresholding [20] for binarization i.e. to convert gray scale images to binary form. 3.3 NORMALIZATION Binary images are normalized to 7x7 size using MATLAB function imresize( )[21]. 3.4 FEATURE EXTRACTION Hindi language is curve based, i.e. each character and modifier is made up of one or more curves, lines and/or loop. In proposed system, feature extraction is done by extracting values of following features for each modifier— i. Horizontal projection profile, ii. Curves’ normalized chain code , and iii. Pixel ratio. Header line or SHIROREKHA plays an important role in formation of Hindi text. Here we use the header line feature for classifying modifiers into different categories. To extract the value of header line feature we apply horizontal projection profile method on normalized modifier image. Then image is thinned by applying [22] and normalized curves’ chain coding scheme and pixel ratio is calculated for each classified modifier to perform final identification & recognition. A. HORIZONTAL PROJECTION PROFILE A projection profile is a histogram of the number of black pixel values accumulated along parallel lines taken through the document [23]. Projection profile methods have been widely and successfully used for skew detection & correction, text line segmentation, word segmentation, page layout segmentation etc. Horizontal projection profile is the method to find the number of character or foreground (black) pixel values collected along rows. Procedure to find horizontal projection profile is given below-


Page 313


ISSN 2319 - 4847

HPP=Procedure Horizontal_Profile(I: grayscale image) Read image I Convert I to binary image B Resize B to the size 64x64 h_proj  zeros for i = 1 to 64 sum  0 for j=1 to 64 if I(i,j)==0 //if black pixel found sum  sum+1 endif endfor h_proj(i)  sum endfor return h_proj On the basis of position of highest profile on x-axis in the range of 0-60, we divided modifiers into three classes— Class

Position of highest profile in HPP (value at x-axis)

Modifier

(BADA AA), 1

Left End (below 20)

(CHOTA U), (BADA U) and (RI) (CHOTI E),

2

Middle (20-40)

(BADI E), (CHOTA AU) (BADA AU)

3

Right End (Above 40)

(CHOTA AE)

(BADA AE) Table1 Classification of modifiers into three classes using horizontal projection profile B. CURVES’ NORMALIZED CHAIN CODE Curves’ normalized chain codes are obtained by following one of the oldest techniques of feature extraction- the chain coding scheme, in computer vision. It was introduced in 1960s. In this scheme, a set of pixels in contour of a shape of an object or character is translated into a set of connections between them. Contour is traced by starting from one pixel and determining one of the 8 adjacent directions in which the next pixel is to be found. Chain code is formed by concatenating the number that designates the direction of the next neighboring pixel as shown in figure. Process is repeated till the starting pixel is reached [24].

Figure 4 Numbering of neighboring pixels in chain coding Proposed method is based on chain coding scheme with a slight difference. In this method, starting point and ending pixels are different. Values are obtained in the range of 0.25 -0.55. Step-by-step method is given in procedure CHN_CODE_FEATURE_EXTR ().


Page 314


ISSN 2319 - 4847

Following variables are used— CP : Current pixel SUM : Stores sum of chain codes of all black pixels in the curve. Initial value is set to zero. I : Counter variable. Stores number of black pixels in the curve. NCC=Procedure CHN_CODE_FEATURE_EXTR (I: thinned image) Step1. Read image I Step2. Convolute 3x3 mask on modifier image row wise from lower left end to upper right end. P is the centre pixel under the mask. Step3. if pixel P is a black pixel if P is an end point, then P  CP endif endif Step4.Search for neighboring pixels in 8-way connectivity and start moving along the curve clockwise as follows— (a) First searching for P2, P3 or P4 neighboring pixel, then P0, P1, P5, P6 or P7 pixels. if pixel P is a modifier pixel, then (i) I  I+1 (ii) SUM  SUM+N (where N is assigned the number according to the direction given in figure 5). (iii) P  CP and repeat. endif (b) if pixel is a header line pixel if pixel is a junction point or end point if all pixels in the curve are searched at least once then stop else go to step (a) endif endif endif Step5. Calculate normalized chain code by the following equation— NC  N return NC Junction point

E.g. Chain code is – 2 2 2 2 2 2 SUM = 12, N= 6 NC= 0.5

CP

Figure 5 Illustration of normalized chain coding procedure of proposed method C. PIXEL RATIO Third feature is the ratio of black pixels to total number of pixels in the image. Values are obtained in the range of 0.25 0.55. PR=Procedure PIXEL_RATIO(I: thinned image, m: number of rows in I, n: number of columns in I)


Page 315

International Journal of Application or Innovation in Engineering & Management (IJAIEM) Web Site: www.ijaiem.org Email: [email protected] Volume 3, Issue 6, June 2014 TP NB

ISSN 2319 - 4847

: :

Total number of pixels in the image, and Number of black pixels in the image. for i: 1 to m for j: 1 to n PR(i,j)  NB(i,j)/ TP(i,j) endfor endfor return PR Finally procedure FEATURE_EXT() is called— [HPP,NCC,PR]= { HPP = call NCC = call PR = call }

procedure FEATURE_EXT(mi: modifier image) procedure HORIZONTAL_PROFILE(mi) procedure CHN_CODE_FEATURE_EXTR (mi) procedure PIXEL_RATIO(mi)

3.5 FUZZY RULE SET CREATION Fuzzy rule set is created by converting the crisp feature values to fuzzy form and forming fuzzy IF-THEN rules. Linguistic variables for inputs and outputs are shown in the table 2. Table 2: Linguistic variables corresponding to input variables Symbol Linguistic Variable Range LE (left end) 0-20 Horizontal Projection Profile (HPP) CL MD (middle) 20-40 RE (right end) 40-60 VL (very low) 0.25-0.30 L (low) 0.30-0.30 Curves’ normalized chain code NC BA (below average) 0.35-0.40 (NCC) A (average) 0.40-0.45 H (high) 0.45-0.50 VH (very high) 0.50-0.55 VL (very low) 0.25-0.30 L (low) 0.30-0.30 Pixel ratio(PR) PR BA (below average) 0.35-0.40 A (average) 0.40-0.45 H (high) 0.45-0.50 VH (very high) 0.50-0.55 Feature

3.5

Total 33 fuzzy rules are created to link these 15 input membership values to 10 output membership values as follows--

R1: IF CL is LE and NC is A and PR is VL THEN MF is

(BADA AA).

R2: IF CL is MD and NC is H and PR is BA THEN MF is

(CHOTI E).

R3: IF CL is MD and NC is H and PR is A THEN MF is

(CHOTI E).

R4: IF CL is MD and NC is VH and PR is BA THEN MF is

CHOTI E).

R5: IF CL is MD and NC is VH and PR is A THEN MF is

(CHOTI E).

R6: IF CL is MD and NC is VL and PR is BA THEN MF is R7: IF CL is MD and NC is VL and PR is A THEN MF is

(BADI E). (BADI E).

R8: IF CL is LE and NC is VL and PR is A THEN MF is

(CHOTA U).

R9: IF CL is LE and NC is VL and PR is H THEN MF is

(CHOTA U).


Page 316


ISSN 2319 - 4847

R10:IF CL is LE and NC is BA and PR is A THEN MF is

CHOTA U).

R11:IF CL is LE and NC is BA and PR is H THEN MF is

(CHOTA U).

R12:IF CL is LE and NC is VL and PR is A THEN MF is

(BADA U).

R13:IF CL is LE and NC is VL and PR is VH THEN MF is R14:IF CL is LE and NC is L and PR is A THEN MF is

(BADA U). (BADA U).

R15:IF CL is LE and NC is L and PR is VH THEN MF is

(BADA U).

R16:IF CL is LE and NC is BA and PR is L THEN MF is

(RI).

R17:IF CL is LE and NC is BA and PR is BA THEN MF is

(RI).

R18:IF CL is LE and NC is A and PR is L THEN MF is R19:IF CL is LE and NC is A and PR is BA THEN MF is R20:IF CL is RE and NC is BA and PR is L THEN MF is

(RI). (RI). (CHOTA AE).

R21:IF CL is RE and NC is BA and PR is BA THEN MF is

(CHOTA AE).

R22:IF CL is RE and NC is VL and PR is A THEN MF is

(BADA AE).

R23:IF CL is RE and NC is VL and PR is H THEN MF is

(BADA AE).

R24:IF CL is RE and NC is BA and PR is A THEN MF is

(BADA AE).

R25:IF CL is RE and NC is BA and PR is H THEN MF is

(BADA AE).

R26:IF CL is MD and NC is H and PR is BA THEN MF is

(CHOTA AU).

R27:IF CL is MD and NC is H and PR is A THEN MF is R28:IF CL is MD and NC is VH and PR is BA THEN MF is R29:IF CL is MD and NC is VH and PR is A THEN MF is R30:IF CL is MD and NC is VL and PR is BA THEN MF is

(CHOTA AU). (CHOTA AU). (CHOTA AU). (BADA AU).

R31:IF CL is MD and NC is VL and PR is A THEN MF is

(BADA AU).

R32:IF CL is MD and NC is L and PR is BA THEN MF is

(BADA AU).

R33:IF CL is MD and NC is L and PR is A THEN MF is

(BADA AU).

Figure 6 Set of fuzzy rules 3.6 CREATION OF PROPOSED NEURO-FUZZY SYSTEM Adaptive Neuro-fuzzy Inference System (ANFIS) is one of the most widely used way of combining both fuzzy set theory and neural networks. Proposed ANFIS consists of six following layers — i. Input node layer – every node in this layer corresponds to input feature used for recognition. ii. Degree of membership layer – nodes in this layer defines the degree of membership of each lingusitic variable corresponding to each input node. iii. Rule node layer – nodes are used to show the IF part of fuzzy rules. Nodes are also known as antecedent nodes. Each node output represents the firing strength of a rule. iv. Normalization layer – also termed as consequent node layer. Each node calculates the ration of the rule’s firing strength to the sum of all rule’s firing strength. Outputs are called normalized firing strength. v. Defuzzification layer – nodes in this layer are adaptive. vi. Output layer – it contains a single aggregate neuron.

4. EXPERIMENTAL RESULTS Volume 3, Issue 6, June 2014

Page 317


ISSN 2319 - 4847

Results of experiment during various stages are as follows— Data collection and scanning— For the experiment 1000 (100 for each) samples of 10 modifiers are collcted from 10 different people of different age groups. A sample of data set is shown below—

Figure 7 Sample data set of modifiers Results of extraction of horizontal projection profile for each modifier are : Modifier

Scanned image sample 1

Horizontal Projection profile of sample 1

Scanned image sample 2

50

Horizontal Projection profile of sample 2 40

45

35

40 30

35 25

30 25

20

20

15

15

10

10 5

5 0

0

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

50

60

45 50

40 35

40

30 30

25 20

20

15 10

10 5

0

0

10

20

30

40

50

60

70

70

0

60

60

50 50

40 40

30

30

20

20 10

10 0

0

10

20

30

40

50

60

70

0

50

70

45

60 40 35

50

30

40

25 20

30 15

20

10 5

10 0

0

10

20

30

40

50

60

70

0

60

60

50

50

40

40

30

30

20

20

10

0

10

0

10

20

30

40

50

60

70

0

0

10

50

40

45

35

40

30

20

30

40

50

60

70

35 25

30 20

25 15

20 10

15 5

10

0

5 0

0

10

20

30

40

50

60

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

0

10

20

30

40

50

60

70

70

70

50

45

60 40

50

35

30

40 25

30

20

15

20 10

10

5

0

0

10

20

30

40

50

60

70

40

0

70

35

60 30

50 25

40

20 15

30 10

20 5

10

0 0

10

20

30

40

50

60

70

0

60

60

50

50

40

40

30

30

20

20

10

0

10

0

10

20

30

40

50

60

70

0

60

40 35

50 30

40 25 20

30

15

20 10

10

5 0

0

10

20

30

40

50

60

70

0

Figure 8 Horizontal projection profiles of selected modifiers System is designed with following parameters—


Page 318

International Journal of Application or Innovation in Engineering & Management (IJAIEM) Web Site: www.ijaiem.org Email: [email protected] Volume 3, Issue 6, June 2014 Number of nodes in input layer Number of nodes in degree of memership layer Number of nodes rule layer Number of nodes in normalization layer Number of nodes in defuzzification layer

: : : : :

ISSN 2319 - 4847

3 15 15 33 10

Structure, FIS editor, plots of membership functions of input variables, rule editor, surface viewer and rule viewer is shown in figure 9 (a)-(e) --

(a)

(b)

(c) (d) (e) Figure 9 (a) Structure of the proposed system, (b) FIS editor (c) Plot of membership function for input variable CL, (d) Plot of membership function for input variable NC and (e) Plot of membership function for input variable PR Rule editor, suface viewer and rule viewer of the proposed system are shown in figure 10 (a)-(c).


Page 319


ISSN 2319 - 4847

(a)

(b)

(c) Figure 10 (a) Rule editor (b) Surface viewer and (c) Rule viewer


Page 320


ISSN 2319 - 4847

To train the neural network in supervised mode, we give linguistic values to each modifier so that the network could train itself to achieve that linguistic value after training number of patterns (of modifiers)Table 3 Linguistic variables corresponding to output variables Modifier

Linguistic Value 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

So inputs and outputs are denoted as: CLASS : μA(CL) where A = { LE,MD,RE} Normalized chain code : μB(NC) where B = {VL,L,BA,A,H,VH} Pixel ratio : μc(PR) where C = {VL,L,BA,A,H,VH} Modifier : μD(MODIFIER) where D = {0.1,0.2,0.3,0.4,0.5,0.6, 0.7,0.8,0.9,1.0} A glimpse of the set of 1000 values is shown below, that are used to train the system-CL

NC

PR

MODIFIER

12

0.45

0.27

0.1

20

0.476

0.288

0.1

5

0.486

0.276

0.1

27

0.459

0.398

0.2

33

0.515

0.365

0.2

58

0.399

0.49

1

56

0.376

0.423

0.8

26

0.299

0.367

0.3

25

0.267

0.442

0.3

34

0.325

0.444

0.9

22

0.255

0.42

0.4

2

0.468

0.213

0.1

23

0.283

0.354

0.3

9

0.313

0.503

0.5

8

0.325

0.515

0.5

27

0.279

0.434

0.3


Page 321


ISSN 2319 - 4847

29

0.285

0.45

0.3

5

0.476

0.299

0.1

30

0.354

0.415

0.9

31

0.324

0.456

1

50

0.389

0.343

0.7

11

0.3

0.5

0.5

51

0.396

0.367

0.7

56

0.266

0.432

0.8

14

0.458

0.278

0.1

22

0.523

0.385

0.2

12

0.401

0.376

0.6

26

0.525

0.367

0.2

53

0.454

0.4

0.8

30

0.467

0.435

0.2

17

0.261

0.432

0.4

45

0.355

0.365

0.7

18

0.267

0.423

0.4

7

0.272

0.514

0.5

11

0.445

0.288

0.1

10

0.322

0.526

0.5

58

0.286

0.426

0.8

6

0.272

0.52

0.5

11

0.356

0.356

0.6

32

0.354

0.463

1

13

0.365

0.383

0.6

14

0.374

0.402

0.6

15

0.406

0.412

0.6

45

0.356

0.312

0.7

12

0.343

0.402

0.6

48

0.362

0.4

0.7

49

0.373

0.352

0.7

38

0.384

0.306

1

57

0.367

0.418

0.8

11


0.301 0.429 Figure 11 A Set of 50 values fed to the system

0.4

Page 322


ISSN 2319 - 4847

After training of the network, recognition rate of each modifier is obtained. Graphical representation of recognition rate of each modifier is shown in figure --

Figure 12 Graphical representation of recognition rate of each modifier

5. COMPARATIVE STUDY A comparative study for the proposed work has been presented in which the proposed work is compared with methods of Sheth et al.[25], Holambe & Thool[26], Kumar & Ravichandran[27] and Nirve & Sable[28]. In this comparative study, we compared proposed system with other neuro-fuzzy systems designed for handwritten characters as well as with systems designed for handwritten Devnagari modifiers. Table 4 Comparison results of proposed method with other methods Technique Dataset Recognition rate Template matching and Gujrathi script modifiers 64% fringe distance classifier Sheth et al. [26] Neuro-fuzzy System Overlapped English 85.92% Characters Holambe & Thool[27] Radial Basis Function Devnagari Consonants 90% Network with modifiers Holambe & Thool [27] Hidden Markov Model Devnagari Consonants 91% with modifiers SureshKumar & Fuzzy Neural Network Handwritten Tamil 90% Ravichandran [28] Characters Nirve & Sable[29] ANFIS Devnagari characters 95% Proposed method ANFIS Devnagari modifiers 97.6% Author Shah & Shrama [25]

Figure 13 Graphical representation of comparative study with other methods


Page 323


ISSN 2319 - 4847

CONCLUSION In this paper, a neuro-fuzzy approach (ANFIS) based classification & recognition system for handwritten Hindi modifiers is presented. Proposed system is also compared with the methods given in [25]-[28] and comparative study shows that the proposed system provides better results than other methods. In this paper, an efficient feature extraction algorithm, which works well with handwritten curve script, is also presented.

REFERENCES [1] Govindan V.K. and Shivprasad A.P., “Character Recognition: A Review ”, Pattern Recognition, vol. 23, no. 7, pp. 671-683, 1990. [2] Haykin S., “Neural Networks : A Comprehensive Foundation”, Pearson Education, 4th Edition, 2013. [3] Yegnanarayana B., “Artificial Neural Networks”, Prentice-Hall of India Pvt. Ltd., Ninth Edition, 2004. [4] Starzyk J. A. and Ansari N., “Feedforward Neural Network for Handwritten Character Recognition”, IEEE symposium on circuit and systems, 1992. [5] Mangal M. and Singh M.P. ,”Handwritten English Vowels Recognition Using Hybrid Evolutionary Feed Forward Neural Networks”, Malaysian Journal of Computer Science, vol. 19(2), pp. 169-187, 2006. [6] Al-Jawfi R., “Handwriting Arabic Character Recognition LeNet Using Neural Network”, The International Arab Journal of Information Technology, vol. 6, no. 3, 2009. [7] Pal A. and Singh D. “Handwritten English Character Recognition Using Neural Network”, International Journal of Computer Science & Communication vol. 1, no. 2, pp. 141-144, 2010. [8] Wu Y., Wu Y., Zhou G. and Wu J., “Recognizing Characters Based on Gaussian-Hermite Moments and BP Neural Networks”, International Conference on Intelligent Computation Technology and Automation, ISBN 978-0-76954077-1, 2010. [9] Zadeh L. A., “ Fuzzy Logic, Neural Networks and Soft Computing”, One page course announcement of CS 294-4, Spring 1993, the University of California at Berkeley, Nov. 1992. [10] Yen J. and Langari R., “ Fuzzy Logic- Intelligence, Control, and Information”, Pearson Education Pub., First Edition , 2006. [11] [11] Alavala C.R. , “ Fuzzy Logic and Neural Networks- Basic Concepts & Applications’, New Age International Pub., First Edition, 2008. [12] Sivanandam S.N. and Deepa S.N., “ Principles of Soft computing”, Wiley India Publication, Second edition, 2012. [13] Sethi I.K. and Chatterjee B. , “ Machine Devnagari Pattern Recognition”, vol. 9, no. 2, pp. 69-75, 1977. [14] Sinha R.M.K. and Mahabala H. , “ Machine Recognition of Devnagari Script, IEEE Trans. System, Man Cybern, 9, pp. 435-441, 1979. [15] Verma B.K. “Handwritten Hindi Character Recognition Using Multilayer Perceptron and: Radial Basis Function Neural Networks”, IEEE International conference on neural networks, vol. 4, pp. 2111-2115, 1995. [16] Kumar S. and Singh C., “ A study of Zernike Moments and its use in Devnagari Handwritten character Recognition”, In Proc. of Int. Conf. on Cognition and recognition, pp. 514-520,2005. [17] Hanmandlu H., Murthy O.V.R. , Madasu V.K. , “ Fuzzy Model Based recognition of Handwritten Hindi Characters”, IEEE Computer Society , Digital Image Computing Techniques and Applications, 2007. [18] Singh G. and Lehri S., “ Recognition of Handwritten Hindi Characters Using Backpropagation Neural Network”, International Journal of Computer Science and Information Technologies, vol. 3(4), 4892-4895,2012. [19] Singh G., Pokhriyal A. and Lehri S., “ Fuzzy Rule Based Classification and Recognition of Handwritten Hindi Curve Script”, International Journal of Computer Engineering & Technology, vol. 4, issue 1, pp. 337-357, 2013. [20] Gonzalez R.C. and Woods R.E. , “ Digital Image Processing”, Pearson Printice Hall Pub., Third Edition, 2008. [21] Image Processing and Neural Network Toolbox help of MATLAb R2009. [22] Pokhriyal A. and Lehri S., “ Minutiae Extraction Using Rotation Invariant Thinning”, International Journal of Engineering Science and Technology, vol. 2(7), pp. 3225-3235, 2010. [23] R. Kasturi, L. O. Gorman, and V. Govindaraju, “Document image analysis: A Primer,” Sadhana Part 1, vol. 27, pp. 3–22, 2002. [24] Nixon M.S.and Aguado A.S. , “ Feature Extraction and Image Processing”, Newness Publication, First Edition, 2002. [25] Shah S.K. and Sharma A. , “ Design and Implementation of Optical Character Recognition System to Recognize Gujarati Script using Template Matching”, IE(I) Journal-Et, vol. 86, pg. 44-49, 2006. [26] Sheth R., Patil K.,Thakur N. and Talele K.T. , “ Neuro –Fuzzy Recognition of Overlapping Handwritten Text Between Adjacent Lines of Text Using Soft Computing”. [27] Holambe A.K. and Thool R.C., “ Comaparative Study of Different Classifier for Devnagari Handwritten Character Recognition”, International Journal of Engineering Science and Technology, vol. 2, no.7, 2681-2689, 2010. [28] Kumar C.S. and Ravichandran T., “ Character Recognition Using RCS with Neural Network”, International Journal of Computer Science Issues, vol. 7, issue 5, pp.289-295, 2010.


Page 324


ISSN 2319 - 4847

[29] Nirve S.A. and Sable G.S., “Optical character Recognition for Printed Text in Devanagari Using ANFIS”, International Journal of Scientific & Engineering Research, vol. 4, issue 10, pp.236-241, 2013.

AUTHOR Gunjan Singh received the degree in Master of Computer Applications from Dr. B.R. Ambedkar University, Agra, UP, India in 2002. She is doing her Ph.D. in Computer Science. Currently, she is assistant professor in MCA department at R.B.S.M.T.C., Agra, UP, India. She has published many papers in several reputed journals such as IJCSIT, IJCET, etc. Her research interests include character recognition, pattern recognition, soft computing, etc.

Dr. Avinash Pokhriyal received the degree in Master of Computer Applications from Dr. B.R. Ambedkar University, Agra, UP, India in 1997. He received the Ph.D. degree in Computer Science on the topic Minutiae based Automatic Fingerprint Recognition. Currently, he is Associate professor in MCA department at R.B.S.M.T.C., Agra, UP, India. He has published and presented several papers in national and international journals and conferences. He has reviewed many papers in several reputed journals such as IJCSNS, JISE, JITE, WASET, IJCSES, etc. He is a member of Editorial board of reputed international journals like IJoFCS, IJCA, IJBB, IACSIT, IJPS, etc. His research interests include biometric based recognition, image processing,pattern recognition, soft computing, etc. Prof. Sushma Lehri received the M.Tech. degree in Computer Science from the DOEACC, India in 2000. She received the Ph.D. degree in Solid State Physics from the Dr. B.R.Ambedkar University, Agra, UP, India in 1980. She has also been the Director of I.E.T., Dr. B. R. Ambedkar University. She has many publications in national & international journals. She has many awards such as research fellowship and postdoctoral fellowship by university grant commission (UGC), New Delhi to her name. Her research interests include image processing, network security, and solid state physics.


Page 325