An Efficient Iris Recognition System Based on ... - Semantic Scholar

9th WSEAS International Conference on NEURAL NETWORKS (NN’08), Sofia, Bulgaria, May 2-4, 2008

An Efficient Iris Recognition System Based on Modular Neural Networks H. ERDINC KOCER, NOVRUZ ALLAHVERDI Electronics and Computer Education Department Selcuk University Technical Education Faculty Alaeddin Keykubad Campus, Konya TURKEY [email protected] http://tef.selcuk.edu.tr Abstract: - In this paper, we propose a neural network based iris recognition approach by analyzing iris patterns. The iris recognition system consists of iris localization, feature extraction and classification of the iris images. Hough transforms were used for localizing the iris region; Cartesian to polar coordinate transform was used for transforming the ring shaped iris image to the rectangular shape. Then, histogram equalization was applied to the iris image for making the shapes in image more distinctive. Average absolute deviation (AAD) algorithm was used for feature extraction in this approach. In matching process, Multi-Layered Perceptron (MLP) and Modular Neural Networks (MNN) are applied to the iris feature vector for classifying the iris images. In fact, this research is focused on measuring the performance of MNN in iris recognition system compared with Multi-Layered Perceptron (MLP) neural network. The gray-level iris images, experimented in this work, were obtained from Institute of Automation Chinese Academy of Science (CASIA) iris images database and Departments of Informatics University of Beira Interior (UBIRIS) iris images database. Experimental results are given in the last stage of this paper. Key-Words: - Biometric, Iris recognition, Modular Neural Networks. An efficient iris recognition system based on ANNs was presented by comparing the performances of MNN and MLP neural networks in this paper. In section 1, the general information about this work is introduced. In section 2, the related works on iris recognition were mentioned. Section 3 has described the preprocessing phase including the iris localization. In section 4, feature extraction method is introduced. Section 5 has presented the neural networks methods including MLP and MNN. Experimental results are presented in Section 6. These results are concluded in section 7

1 Introduction Biometric recognition is a very effective technique using for recognizing the identity of a person. This recognition method is based on human physiological characteristics such as, fingerprint, facial patterns, iris texture, hand geometry etc. and behavioral characteristics such as voice print or speech pattern, handwritten signature etc. Among various biometric techniques, iris recognition has become most preferred method in the last decade. Because iris features are stable over lifetime and distinctive. This allows more reliability in identification and verification applications. Also, compared with other biometrics such as fingerprint, hand geometry etc., iris recognition doesn’t contain physical contact during scanning the input pattern [1, 4, 11, 12, 16]. In most of the iris recognition systems, the matching process is based on statistical methods such as Hamming distance, Euclidian distance etc. In the statistical method, the matching process is work as comparison of two images one-to-one. This type of comparison has some disadvantages. If the matching images are not clear or these images contain lots of noises such as reflections, eyelids, eyelashes etc., these images can’t be recognized properly. On the contrary, artificial neural networks (ANNs) method can cope with these kinds of problems, because this method is based on training of the feature vector of the input images. These input images may include noisy or blurred images, so the recognition accuracy will be increased.

ISBN: 978-960-6766-56-5

2 Related Works John Daugman [4], scientist in Cambridge University, developed very efficient method for iris recognition in 1992. In Daugman’s system, Integro-differential operator was used for detecting the iris boundaries and 2D Gabor filter was used for feature extraction. He generated 256 byte iris code and used XOR operator for comparison of the iris images. This operator analyses the Hamming Distance, between iris codes by making bit to bit test. Reillo et al. [12], Tisse et al. [1], Masek [8] and Sun et al. [20] also used Hamming Distance method for matching the iris images. Richard P. Wildes [11] described a system for iris recognition in 1997. He used Hough Transform for locating the iris and generated iris code by using Laplacian Pyramid (also called multiscale decomposition). Wildes developed Normalized Correlation algorithm for matching process. W.W. Boles

42

ISSN: 1790-5109


and Boashah [17] proposed dissimilarity function for iris recognition. Ma et al. [7], Huang et al. [19] and Noh et al. [14] used Euclidian Distance in matching process. In recent years, artificial neural network based iris recognition systems were chosen rather frequently. Lim et al. [13] were used LVQ (Learning Vector Quantization) algorithm for classification of iris images in 2001. Liam et al. [6] were classified iris images by using SOM (Self Organizing Map) algorithm. Alim et al. [10] were used MLP for recognizing the iris images.

3.2 Cartesian to Polar Coordinate Transform The Cartesian to Polar Coordinate Transform, suggested by the J. Daugman [4], allows us to transform the circular images to the rectangular shaped images with fixed dimensions. This method remaps each point within the iris region to the polar coordinates (r, θ), where r is the radius calculated as subtraction of inner circle radius from outer circle radius, and θ is an angle. The transform equation can be defined as: I (x(r,θ) , y(r,θ)) => I (r, θ) Here, x(r,θ) and y(r,θ) can be found in:

3 Pre-Processing of the Iris Images

(3)

x(r,θ) = (1-r) xp(θ) + rxi (θ) y(r,θ)= (1-r) yp(θ) + ryi (θ) (4) where I(x,y) represents the iris image, xp, yp and xi, yi are the coordinates of the pupil and iris boundaries. We have obtained the locating 140x70 pixel iris images for CASIA database [9] and 80x40 pixel iris images for UBIRIS database [5]. The sample images taken from CASIA and UBIRIS iris image database were shown in Figure 1.a and b; the eye images including the inner and outer boundaries marked with white line shown in Figure 1.c and d; the segmented iris region, marked with upper and lower black rectangles, were shown in Figure 1.e and f; the iris region images, obtained from Cartesian to polar coordinate transform, were shown in Figure 1.g and h.

The pre-processing of iris images includes iris localization. Iris localization process is very critical, because it affects the success rate of iris recognition system directly. If the iris region couldn’t be obtained properly, the iris image would not be classified. Since captured eye image contains some unused parts such as eyelids, pupil etc., so that the inner and outer boundaries of the iris region should be determined. We employed Hough transform method for determining the boundaries of the iris. Then, Cartesian to polar coordinate transform method was used for converting the ring shaped iris pattern to the rectangular form. Histogram equalization technique is employed to make the iris patterns more distinctive before feature extraction.

3.1 Hough Transform The Hough transform is designed to find lines, since a line can be defined as a collection of edge points. This transform is an algorithm that will take a collection of edge points as found by an edge detector. If we define the line trigonometrically [4, 17]:

(a)

(c)

(e)

(g)

(b)

(d)

(f)

(h)

X sinθ + Y cosθ = r (1) where (x,y) denotes coordinates in the spatial domain and (r,θ) denotes the transform parameters. The circular Hough transform, employed by Wildes et al. [11], can be used for detection of the circular lines. The equation for circular Hough transform can be defined as:

3.3 Histogram Equalization

Xc2 + Yc2 = r2 (2) where (Xc,Yc) are center coordinates of a circle and r is the radius. A maximum point in the Hough space will correspond to the radius and center coordinates of the circle best defined by the edge points After getting the iris boundaries, the noisy regions such as eye lids, eyelashes etc. should be eliminated. So, we have defined imaginary horizontal lines from the centre of the pupil to the lower direction of the iris image with the size of predetermined pixel number from the centre point of the pupil.

The gray-level histogram of an image is the distribution of the gray level values in an image. The histogram equalization is a popular technique to improve the appearance of a poor contrasted image. The process of equalizing the histogram of an image consists of 4 steps [15]: (1) Find the sum of the histogram values. (2) Normalize these values dividing by the total number of pixels. (3) Multiply these normalized values by the maximum gray-level value. (4) Map the new gray level values. The sample of equalized iris images taken from CASIA and UBIRIS databases were shown in Fig. 2.a

ISBN: 978-960-6766-56-5

Fig. 1 Iris localization and segmentation process

43

ISSN: 1790-5109


and 2.b respectively. As it can be seen in these figures, the distinguished features of the iris images have become clearer.

(a)

5 Classification of the Iris Images The last step of the iris recognition system is comparison of the iris templates. Therefore the authorization process can be made. In the proposed approach, two ANN models were used for classification. These are MLP and MNN. For training the network, back-propagation algorithm was chosen. Almost %70 of neural network studies have used back-propagation algorithm for training process. For measuring the training performance of the network, mean square error (MSE) function is used. The value of the MSE is used to determine how well the network output fits the desired output. The stop criteria for supervised training are usually based on MSE. Most often the training is set to terminate when the MSE drops to some threshold. The formula for calculating the MSE value is;

(b)

Fig. 2 Histogram equalized iris images

4 Feature Extraction Feature extraction provides us to obtain the most discriminating information of an image. This information can be presented as a feature vector. A feature vector that includes global and local features of an iris image should be encoded so that the comparison between iris templates can be made [15]. In the proposed approach, the feature vector of an iris image was encoded by using Average Absolute Deviation algorithm [7]. This algorithm is defined as:

V =

1⎛ ⎞ ⎜ ∑ f ( x, y ) − m ⎟ N⎝ N ⎠

MSE =

(5)

5.1 Multi-Layered Perceptron The The processing units in MLP are arranged in three layers. These are input layer (includes the information you would use to make decision), hidden layer (helps network to compute more complicated associations) and output layer (includes the resulting decision) [18]. Each neuron in the input layer is fed directly to the hidden layer neurons via a series of weights. The sum of the products of the weights and the inputs is calculated in each node. The calculated values are fed directly to the output layer neurons via a series of weights. As in hidden layer, the sum of the products of the weights and the hidden layer neuron outputs is calculated in each node in the output layer. If the error between calculated output value and the desired value is more than the error ratio, then the training (changing the weights and calculating the new output by using the new weights) process begins. This training process can be finished by obtaining the desired error rate for all input combinations. The structure of basic feed-forward MLP neural network is shown in Figure 4.

AAD 0.633851

…...

(6)

Here, y is the desired output, y’ is the output value calculated by the system, N is the number of output node. Approaching the MSE value to the zero means that the calculated output value is becoming the closer to the desired output value.

where N is the number of pixels in the image, m is the mean of the image and f(x,y) is the value at point (x,y). In our work, we divided the iris images by the various dimension of sub-images. The 140x70 pixel images (CASIA iris images) were divided into 10x10 and 14x14 pixel sub-images (Fig. 3). The 80x40 pixel images (UBIRIS iris images) were divided into 10x10 and 20x20 pixel sub-images. Each sub-image was encoded by using AAD algorithm. The feature extraction process is shown in figure 3. We obtained the feature vectors with the length of 98 byte and 50 byte for the 140x70 pixel iris images. For 80x40 pixel iris images, the feature vectors have a length of 32 byte and 8 byte. The entire feature vectors were applied to the Neural Network as an input for classification of the iris images.

0.001455

1 N ' ( yi − yi ) 2 ∑ N i =1

0.813345

Fig. 3 Histogram equalized iris images

ISBN: 978-960-6766-56-5

44

ISSN: 1790-5109


b

Where n is the presentation number and α is constant that determines the effect of previous weight changes on the current direction of movement in the weight space. For a network to convergence to the required solution it is desirable to use higher learning rate (η).

b j1 Σ F

j2

k1 o1

t1

Σ F

o2

t2

Target vector

Σ F

Activation vector

Σ F

x2

Output layer

x1

kp W ij

W jk

5.2 Modular Neural Networks A modular neural networks (MNN), which is developed at the beginning of nineties, is the specialized class of MLP neural network [2]. The processing units in MNN are arranged in three layers as in MLP. But the hidden layer is composed by the modules of hidden layer. Most of MNN structure has two or more hidden layers, but in recent years, the MNN structure has been constructed as combination of many MLP modules. In other words, MNN is developed as many modules and these modules are included miniaturized MLP networks. An example of this type of this type MNN is shown in Figure 5.

jm Σ F

Input layer

Hidden layer

xn

Fig. 4 A basic MLP neural network structure Let the net input to a node in layer j be given by

X1

n

net j = ∑ w ji oi

(7)

MLP

i =1

The output of node j be

MLP

INPUTS

o j = f ( net i ).

(8) Here f is the activation function; in this study a hyperbolic tangent function has been used

MLP MLP

2 ( net +θ ) / θ

e j j 0 −1 o j = 2 ( net j +θ j ) / θ 0 e +1

Fig. 5 The structure of MNN (10) In this approach, a network that is comprised of several subcomponents is used to solve a task. The structures of these subcomponents can vary greatly, and their number can be dynamic. Consequently, every component of the neural network has the potential of specializing in a subtask; thus, this architecture is more powerful than a standard architecture [3]. The structure of MNN used in this research is shown in Figure 6.

ok = f ( net i ).

(11) For each input-output pattern, the square of the error can be given by

1 (t k − o k ) 2 ∑ 2 k

MLP

XN

Its respective outputs are

E=

OUTPUT

MLP

(9)

Now, the input to the nodes of layer k is

net k = ∑ wkj o j

MLP

(12)

For any output layer k, ΔWkj will be given by

Δ wkj = η (t k − o k ) f k ( net k ) oi = ηδ k o j

(13)

Similarly for the internal units

Δ w ji = ηδ j oi

(14) One way to increase the learning rate without leading to oscillation is to modify the back propagation algorithm by including the momentum term α as below:

Δw ji ( n + 1) = η (δ j oi ) + α Δw ji ( n ).

ISBN: 978-960-6766-56-5

OUTPUTS

INPUTS

Hidden Layers

Fig 6 The structure of MNN composed by two hidden layers

(15)

45

ISSN: 1790-5109


Here, the structure of the network (number of neurons in each hidden layer) is determined by using heuristic searching methods same as MLP network.

is 100 % and is obtained in configuration of 10x10 subimaged. The computer used for training process has P4 3 GHz microprocessor and 1024 Mb DDR2 memory hardware.

6 Experimental Results 7 Conclusion

In order to evaluate the performance of the proposed system, 490 iris images (captured from 70 people, 7 images per person) selected from CASIA database and 300 iris images (captured from 60 people, 5 images per person) selected from UBIRIS database were employed. The dimension of CASIA iris images is 320x280 pixels and the dimension of UBIRIS iris images is 200x150 pixels. Sigmoid transfer function is used for activation of neurons. Quick back propagation learning algorithm was used for training the ANN. Maximum 10000 iterations were performed for each input set. Only one of the input images was used for testing the system, the rest was performed in training the system. The testing results were given in Table 1 and Table 2.

In this paper, an efficient iris recognition system based on ANNs was presented. The proposed system was employed Hough Transform and Cartesian to Polar Coordinate Transform for iris localization, Histogram Equalization and Average Absolute Deviation for feature extraction. ANNs with Back-propagation learning algorithm was implemented for classifying the iris images. Multi-layered perceptron and modular neural network was performed comparatively for recognition of human iris images. The correct classification rates are in acceptable limit in recognition systems. Also, it has been observed that the disadvantages of statistical methods have been annihilated. Experimental results were shown that the MNN is more efficient than MLP in iris recognition system.

Table 1 Experimental Result for MLP network Iris Database & Dimension of SubImages

References: [1] C.L. Tisse, L. Martin, L. Torres and M. Robert. “Person identification technique using human iris recognition”, The 15th International Conference on Vision Interface, 2002, pp.294–299. [2] D.W. Patterson, “Artificial Neural Networks. Theory and Applications”, Prentice Hall, 1996. [3] E. Schlessinger, P. Bentley and R. Lotto, “Modular Thinking: Evolving Modular Neural Networks for Visual Guidance of Agents”, Genetic and Evolutionary Computation Conference, USA, 2006, pp 215-222. [4] J. Daugman, “High Confidence Visual Recognition of Persons by a Test of Statistical Independence” IEEE Transactions on Pattern Analysis and Machine Intelligence, 15, No.11, 1993, pp. 1148-1161. [5] H. Proenca, L. Alexandre, “UBIRIS: A noisy iris image database”, Proceedings of Int. Conference on Image Analysis and Processing, Vol.1, 2005, pp. 970-977. [6] L. Liam, A. Chekima, L. Fan and J. Dargham. “Iris recognition using self-organizing neural network”, IEEE, 2002 Student Conference on Research and Developing Systems, 2002, pp. 169–172. [7] L. Ma, Y. Wang, T. Tan, “Iris recognition based on Multichannel Gabor Filtering”, The 5th Asian Conference on Computer Vision, Australia, 2002. [8] L. Masek, “Recognition of Human Iris Patterns for Biometric Identification”, Bachelor of Eng. Thesis, the University of Western Australia, 2003. [9]NLPR 2007.

Correct Classification Rate – CCR (%) MLP Iteration

MSE

Duration (min : sec)

Testing

CASIA10x10

7110

0,0029

14 : 32

94,28

CASIA14x14

7385

0,0031

11 : 28

94,28

UBIRIS10x10

7015

0,0024

02 : 36

98,33

UBIRIS20x20

9687

0,0107

04 : 25

88,33

Table 2 Experimental Result for MNN network Iris Database & Dimension of SubImages

Correct Classification Rate – CCR (%) MNN Iteration

MSE

Duration (min : sec)

Testing

CASIA10x10

6052

0,0047

11 : 09

97,14

CASIA14x14

3725

0,0069

04 : 17

97,14

UBIRIS10x10

7321

0,0008

08 : 45

100

UBIRIS20x20

5107

0,0082

03 : 49

90,00

As it can be seen from the table, modular neural network has the best testing results. The best CCR for CASIA database is 97.14 % and is obtained in both of 10x10 and 14x14 sub-imaged; for UBIRIS database, the best CCR

ISBN: 978-960-6766-56-5

46

ISSN: 1790-5109


http://nlpr-web.ia.ac.cn/English/irds/resources.htm (15.01.2008). [10] O. A. Alim, and M. Sharkas. “Texture classification of the human iris recognition using artificial neural Networks”, Proceedings of IEEE, Cairo 2002. [11] R.P. Wildes, “Iris Recognition: An Emerging Biometric Technology”, Proceedings of the IEEE, 85, No. 9, 1997, pp. 1348-1363. [12] R.S. Reillo, and C.S. Avila, “Processing of the human iris pattern for biometric identification”, Proceedings of IPMU, 2000, pp 653-656. [13] S. Lim, K. Lee, O. Byeon and T. Kim. “Efficient Iris Recognition through Improvement of Feature Vector and Classifier”, ETRI Journal, 23, No. 2, 2001, pp. 61-70. [14] S. Noh, K. Bae and J. Kim. “A novel method to extract features for iris recognition system”, Proceedings of 4th International Conference on Audio- and Video-Based Biometric Person Authentication, 2003, pp. 838-844. [15] S. Umbaugh, “Computer vision and Image processing”, Prentice Hall, 1999. [16] W. Chen, S. Yuan, “A Novel Personal Biometric Authentication Technique Using Human Iris based on Fractal Dimension Features”, Proceedings of Int. Conference on Acoustics, Speech and Signal Processing, 2003, pp. 201-204. [17] W.W. Boles, B. Boashah, “A human identification Technique Using Images of the Iris and Wavelet Transform”, IEEE Transactions on Signal Processing, 46, 1998, pp. 1185-1188. [18] Y. Ozbay, B. Karlik, “A fast training back propagation algorithm on windows”, Proceedings of 3rd Int. Symp. Mathematical & Computational Application, 2002, pp 204-210. [19] Y.P. Huang, S.W. Luo and E.Y. Chen. “An efficient iris recognition system”. Proceedings of 1st Int. Conference on Machine Learning and Cybermetics, 2002, pp. 450-454. [20] Z. Sun, Y. Wang, T. Tan and J. Cui, “Improving iris recognition accuracy via cascaded classifiers”, IEEE Transactions on Systems, Man. and Cybernetics, 35, no. 3, 2005, pp. 435-441.

ISBN: 978-960-6766-56-5

47

ISSN: 1790-5109