Pp6

ICSP2006 Proceedings

Palmprint Identification using Gabor Wavelet Probabilistic Neural Networks 1 Dongmei Sun, Zhengding Qiu, Qiang Li Institute of Information Science, Beijing Jiaotong University, Beijing 100044, P.R.China E-mail: [email protected] Abstract: In this paper, a novel algorithm for automatic palmprint identification based on Gabor wavelet probabilistic neural networks (GWPNN) is proposed, which incorporating advantages of Gabor wavelet network (GWN) and probabilistic neural networks (PNN). We first construct a GWN model for each individual palmprint and then combine it with PNN in a unified framework that aims to significantly improve the discriminant capability for palmprint recognition. In our experiments, the accuracy identification rate can reach 99.5% on the database that contains 1,971 image samples. It demonstrates that proposed algorithm is effective and robust. Key words: palmprint identification, Gabor wavelet, probabilistic neural networks

This paper proposes a new method for palmprint identification called Gabor wavelet probabilistic neural networks (GWPNN), which incorporating advantages of both Gabor wavelet network (GWN) and probabilistic neural networks (PNN). We first construct a GWN model for each palmprint and then combine it with PNN in a unified framework that aims to significantly improve the discriminability for palmprint identification. The rest of the paper is organized as follows. GWN is described in Section 2. GWPNN is discussed in Section 3. Section 4 gives some experimental results. Finally, conclusions are presented in Section 5.

1. INTRODUCTION

2. GABOR WAVELET NETWORK

Recently, palmprint recognition has received much more attention. Compared with the other available biometric features, palmprint recognition has several advantages [1]: (i) palmprints contain more rich information than fingerprints and hand geometry, so they are more distinctive; (ii) palmprints are easily captured even with lower resolution devices, which would be cheaper; (iii) using a palmprint capturer, all the features of a palm such as hand geometry, minutiae features, principal and wrinkles could be combined to build a more accurate and robust multi-modal biometric system; (iv) user’s acceptability is high. Current works on palmprint recognition primarily are based on two categories of features [2]: (i) statistical features such as eigenpalm [3], fisherpalm [4], Fourier transform [5], Gabor filter [6], wavelet transform [7]; (ii) structural features including principal lines, creases, delta points, minutiae [8-10]. One of important problems for palmprint recognition is how to match or classify a palmprint using the extracted features. There are usually two categories of discrimination techniques for palmprint recognition. One is based on minimum distance classifiers. Some popular distance measurements include Euclidean, Mahalanobis and cosine distance. The other is constructing decision classifiers by optimizing an error criterion. One of examples is artificial neural network.

For an automatic pattern identification system, it is important to accurately extract local features. Wavelet analysis is desirable for this purpose since it has good characteristics of space-frequency localization. In particular, Gabor wavelet functions provide the best possible tradeoff between spatial and frequency resolution [11]. Furthermore, using Gabor wavelets analysis is biologically motivated, as they have similar shapes to the receptive fields of simple cells in the human visual cortex. Wavelet networks are the combination of wavelet theory and neural networks, which using wavelets as activation function. The strength of wavelet networks lies in their capabilities of catching essential features in “frequency-rich” signals. Kruger has proposed a GWN for face recognition [11]. A set of weighted Gabor wavelets were used to represent an image. Since GWN was optimized for the representation of an image, it may not effectively handle the discriminating features for recognition [12]. So we extend the GWN to a new GWPNN. For a given palmprint image f(x,y), a GWN model is constructed by a set of Gabor wavelets ^\ i ` :

f ( x, y )

N

¦ w\ i

i

< 7W

i 1

1

This work is supported by Research Foundation of Beijing Jiaotong University under contracts 2004SM008, the Key Laboratory of Information Science & Engineering of Railway Ministry under Grant No.0509.

____________________________________ 0-7803-9737-1/06/$20.00 ©2006 IEEE

Authorized licensed use limited to: Jerusalem College of Engineering. Downloaded on July 28, 2009 at 23:11 from IEEE Xplore. Restrictions apply.

(1)

(a) Original palmprint images

(b) The representation with GWN for (a)

Fig.1 The palmprint images represented with the optimized GWN X11

\

X12 1

X1N1 D1(x )

\2 Xi1

C(x)

Di(x)

Xi2

Input palmprint

\i XiN2 Dm(x)

Xm1 Xm2

\N

XmNm

Gabor wavelets

Summation layer

Pattern layer

Decision layer

Fig. 2 Diagram of GWPNN Here

^\ i ` is a 2D Gabor wavelet basis function that can

be defined as:

\ i ( x, y)

1 2SV 2

1

ª º exp « 2 ( x'2  y '2 ) » cos(Z0 x' ) 2 V ¬ ¼

(2)

Where,

x'

ai ( x cos T i  y sin T i ), y '

Z0 , V , ai ,Ti

ai ( x sin T i  y cos T i ),

are the frequency, the deviation related to the

spatial band width, the scale and the orientation respectively. In order to find the GWN for an image f, we minimize the energy function:

W

N

arg min f ( x, y )  ¦ wi\ i ( x, y )

2

(3)

Where,

W

^w1 , w2 ,!, wN ` denotes

a weight vector

specific for an individual palmprint image f. N is the number of Gabor wavelets in GWN. Because Gabor wavelets are non-orthogonal, for a given family < , the weight W can not be calculated by projection of the Gabor wavelets onto the image f. According to Kruger [11], this problem can be solved by

 considering the bi-orthogonal family of wavelets < follow: W

7f <

< < 7

1

<7 f

as (4)

Fig.1 shows two original palmprint images represented with the optimized GWN.

i 1

Authorized licensed use limited to: Jerusalem College of Engineering. Downloaded on July 28, 2009 at 23:11 from IEEE Xplore. Restrictions apply.

3. GABOR WAVELET PROBABILISTIC NEURAL NETWORKS

g ij ( x )

ª ( x  x ij ) T ( x  x ij ) º p (i ) exp «  » 2V 2 ¬« ¼»

p(i )

1 Ni

Ni

(5)

T

ª ( x  xij ) ( x  xij ) º » (6) 2V 2 ¬« ¼»

Where, Ni is the number of samples in class Ci. The final decision output is presented at the last layer. The output layer unit classifies the pattern x according to the Bayes’s decision rule based on the output of all the summation layer neurons:

Cˆ ( x)

arg max{Di ( x)}, i 1,2,..., m

Fig.3 The captured palmprint image and the ROI of the palmprint

20 17.5 15

12.5

7.5 5

¦ exp « j 1

(b) The ROI of the palmprint image

10

Where, p(i) is a priori probability, xij is the neuron vector, ı is a smoothing parameter. The forth layer is summation layer, the neurons compute the maximum likelihood of pattern x being classified into Ci through summarizing and averaging the output of all neurons that belong to the same class:

Di ( x)

(a) The original captured palmprint image

Distance

The PNN is one of implementation on Bayes strategy through finding the minimum risk cost based on the probability density function. PNN offers several strengths over back-propagation networks. Besides its generalization ability, the training speed of PNN is much faster because its leaning rule is simple and requires only a single pass through the training data [2]. Most important, training a new pattern into a trained PNN requires no retraining of the existing network links. In this paper, we propose a GWPNN by introducing the weight vectors of GWN to PNN as feature patterns. As show in Fig.2, the diagram consists of five layers. The first layer receives input palmprint images. The second layer corresponds to a set of Gabor wavelets, which are used to represent the image patterns with GWN feature space. The third layer servers as pattern layer, which receives the weight vectors of GWN. On receiving a pattern x from the second layer, the neuron xij of the pattern layer computes its output [13]:

(7)

In this way, GWPNN inherits the strengths of both GWN and PNN in image representation and in pattern classification. 4. EXPERIMENTS The experimental database is consisted of 1,971 righthand images from 98 individuals. To evaluate the stability of palmprint features, each individual’s hand-image is collected 2~5 times in a period of 3 month. A special digital-camera-based device is designed to capture the images. The original resolution of hand-images is 1792×1200. We take the ROIs (region of interest) out and

0

2

4

6

8

10 12 14 16 18 20

User No.

Fig. 4. Distance between a test sample and other 20 samples unify their size to 128×128. After histogram equalization and noise canceling of the ROIs, the palmprint database is set up. Fig.3 shows the original captured palmprint image and the ROI of the palmprint through preprocessing. The hold-out rule is employed to evaluate the classification error. Palmprint database is divided into two non-overlapping sets: training set and testing set. The training set is made up of 490 palmprint (5 palmprints from each of unique individual) samples, while the remainder 1481 images make up of the testing set (6-20 samples for each person). We investigate the identification performance using different number of Gabor wavelets for GWPNN. Table 1 displays the correct identification rate as the number varying from 16 to 100. It is demonstrated that as the number of Gabor wavelets increase, the correct identification rate also increase. When 80 Gabor wavelets are used the identification rate is highest. Even the number of Gabor wavelets increase, the performance did not increase anymore. Fig. 4 shows the distances between a test sample and 20 palmprints in the database. We can see that the output value of the GWPNN is 3#. It is noticeable that using GWPNN we can obtain the correct identification rate 99.5%.

Authorized licensed use limited to: Jerusalem College of Engineering. Downloaded on July 28, 2009 at 23:11 from IEEE Xplore. Restrictions apply.

Table 1. Correct identification rate using different number of Gabor wavelets in GWPNN Number of wavelets Identification rate (%)

16

25

36

75.2 80.5 88.6

64

80

100

96

99.5

98

5. CONCLUSION This paper proposes a new method for palmprint identification called GWPNN, which incorporating advantages of both GWN and PNN. First, we construct a GWN model for individual, then the weight vectors of GWN are introduced to PNN as feature patterns. In this way, a GWPNN not only has good image representation ability, but also significantly improve the discriminating capability for palmprint identification. The experimental results demonstrate the effectiveness and accuracy of the method. REFERENCES [1] A.K. Jain, A. Ross, D. Prabhakar, “An introduction to biometric recognition”, IEEE Trans. Circuits and Systems for Video Technology, vol. 14, pp. 4-20, 2004. [2] T. Connie, A. Jin, M. Ong, D. Ling, “An automated palmprint recognition system”, Image and Vision Computing, vol. 23, pp. 501-515, 2005.

[3] G. Lu, D. Zhang, K. Wang, “Palmprint recongnition using eigenpalm features”, Pattern recognition Letters, vol. 24, pp. 1463-1467, 2003. [4] X. Wu, G. Lu, D. Zhang, K. Wang, “Fisherpalms based palmprint recongnition”, Pattern recognition Letters, vol. 24, pp. 2829-2838, 2003 [5] W. Li, D. Zhang, Z. Xu, “Palmprint Identification by Fourier Transform”, International Journal of Pattern Recognition and Artificial Intelligence, vol. 16, pp. 417-432, 2002 [6] W. Kong, D. Zhang, W. Li, “Palmprint feature extraction using 2-D Gabor filters”, Pattern recognition, vol. 36, pp. 23392347, 2003 [7] L. Zhang, D. Zhang, “Characterization of palmprints by wavelet signatures via directional context modeling”, IEEE Trans. Man and Cybernetics, Part B, vol. 34, pp. 1335-1347, 2004. [8] D. Zhang and W. Shu, “Two novel characteristics in palmprint verification: datum point invariance and line feature matching”, Pattern Recognition, vol. 32, pp. 692-702, 1999. [9] J. Chen, C. Zhang, G. Rong, “Palmprint Recognition using Crease”, International Conference on Image Processing, vol.3, pp. 234-237, Oct. 7-10, 2001. [10] C. Han, H. Cheng, C. Lin, K. Fan, “Personal authentication using palm-print features”, Pattern Recognition, vol. 36, pp. 371381, 2003. [11] V. Krüger, “Gabor Wavelet Networks for Object Representation”, Ph.D thesis, University of Maryland, 2001. [12] H. Zhang, B. Zhang, W. Huang, Q. Tian, “Gabor Wavelet Association Memory for face Recognition”. IEEE Trans. Neural Networks, vol. 16, pp. 275-278, 2005 [13] K. Mao, K. Tan, W. Ser, “Probabilistic Neural-Network Structure Determination for Pattern Classification”, IEEE Trans. Neural Networks, vol. 11, pp. 1009-1016, 2000.

Authorized licensed use limited to: Jerusalem College of Engineering. Downloaded on July 28, 2009 at 23:11 from IEEE Xplore. Restrictions apply.