2009 Sixth International Conference on Information Technology: New Generations

An Authentication System Based on Palmprint Hanmandlu Madasu1, H. M.Gupta1, Neha Mittal2, and Shantaram Vasikarla3
Department of Electrical Engineering, I.I.T. Delhi, India Department of ECE, Vidya College of Engg., Vidya Knowledge Park, Meerut, UP, India 3 American InterContinental University, 12655 W. Jefferson Bl., Los Angeles, CA 90066, USA
2 1

Corresponding Email: mhmandlu@ee.iitd.ac.in

Abstract
In this paper, palmprint based authentication is presented. The palmprint image is acquired using an acquisition system developed at IIT Delhi. The Region of interest (ROI) is extracted from the palmprint image by finding a tangent of curves between fingers. The perpendicular bisector of this tangent divides the rectangular area enclosing the palmprint into two equal parts. The features extracted from the ROIs are used for matching. Two approaches are suggested for the feature extraction. In the first approach the ROI is divided into a suitable number of non-overlapping windows from which fuzzy features are extracted. In the second approach multi-scale wavelet decomposition is applied on the ROI and the detail images are combined to yield a superimposed image which is partitioned into non-overlapping windows. From these windows energy feature is extracted. The two sets features are used to determine the genuine and impostor scores. The results on 125 users show 99.2% with fuzzy feature and 94.4% with wavelet based feature. Keywords: Fuzzy features, wavelet features, Genuine score, impostor scores, ROC.

transform is adopted to decompose a palmprint image into lower and higher frequency subbands. This decomposition reduces the computational load of moment calculation drastically. The generated waveletmoment based feature representation is used to create cancelable verification key with a set of random data. This private binary key can be canceled and replaced. Besides that, this key also possesses high data capture offset tolerance, with highly correlated bit strings for intra-class population. This property allows a clear separation of the genuine and imposter populations, as well as zero Equal Error Rate attainment, which is hardly gained in the conventional biometric based authentication system. Matching of palm prints in [5] deals with the feasibility of person identification based on feature points extracted from palmprint images. In this approach, a set of feature points is extracted along the prominent palm lines (and the associated line orientation) from a given palmprint image. Next it is decided whether two palmprints belong to the same hand by computing a matching score between the corresponding sets of feature points of the two palmprints. The two sets of feature points/orientations are matched using point matching technique which takes into account the nonlinear deformations as well as the outliers present in the two sets. The estimates of the matching score distributions for the genuine and imposter sets of palm pairs indicate that palmprints have a good discrimination power. Palmprint Verification with moments in [8] introduces an experimental evaluation of the effectiveness of utilizing three well known orthogonal moments, namely Zernike moments, pseudo Zernike moments and Legendre moments in the application of palmprint verification. The idea of implementing orthogonal moments as palmprint feature extractors is prompted by the fact that the principal features of palmprint are based on the line structure. These orthogonal moments are able to define statistical and geometrical features containing line structure information about palmprint. Experimental results show that the performance of the

1. Introduction
The palmprint as a biometric modality is recently recognized. Slowly but decisively it is gaining acceptance in biometrics. As compared to other biometric modalities, it is bestowed with enormous information which can be used as the discriminating power. We will discuss a few important contributions made on this modality. A palmprint authentication system in [7] is specifically designed to overcome the limitations of the contemporary biometric authentication systems. In this system, Geometric and pseudo Zernike moments are employed as feature extractors to transform palmprint image into a lower dimensional compact feature representation. Before moment computation, wavelet
978-0-7695-3596-8/09 $25.00 2009 IEEE DOI 10.1109/ITNG.2009.297 399

system is dependent on the moment order as well as the type of moments. The orthogonal property of these moments is able to characterize independent features of the palmprint image and thus have minimum information redundancy in a moment set. Palmprint verification using complex wavelet transform [9] moots a modified complex wavelet structural similarity index (CW-SSIM) as the matching score for identifying the input palmprint. The local structure information of a palm print is hidden in the relative phase patterns of the complex wavelet coefficients and a constant phase shift of all coefficients does not change the structure of local image features. Since CW-SSIM is robust to translation, small rotation and distortion, a fast rough alignment of palmprint images is sufficient. CWSSIM is also insensitive to luminance and contrast changes. Palmprint classification in [10] serves as an important indexing mechanism in a palmprint database. This algorithm uses a novel representation and is based on a two-stage classifier. The two-stage classification algorithm gives even-distributed categories. A novel representation scheme is directly derived from the principal line structures. The representation does not use wrinkles, and singular points. It is capable of tolerating poor image quality. In this paper, we will explore both fuzzy features and wavelet features extracted from a palmprint. Though Gabor features have been extensively used on palmprints, it is observed that they are sensitive to slight changes in size and orientation of a palm. An effort is made to work on two counts: An efficient extraction of Region Of Interest (ROI) and an effective feature selection. The organization of the paper is as follows: Section 2 presents the extraction of ROI. The wavelet features are described in Section 3 whereas fuzzy features in Section 4. Results of implementation are described in Section 5 and finally conclusions are drawn in Section 6.

proposed ROI extraction are inspired from the work of [6]. 2.2 Preprocessing method for ROI extraction The following steps outline the ROI extraction from a palm: 1. 2. 3. 4. 5. Take the original image (See Fig. 1) and convert it from RGB to grayscale. Crop a fixed section of the image not having the glass in it (See Fig. 2). Rotate this section based on the type of hand - left or right so that the image is in a uniform direction Take the image and find its histogram. Take the moving average of the histogram, and find the minimum of this moving average. The point of minimum provides us with the threshold level to be used to convert the image into a binary image. The binary conversion [2] is done inversely, i.e. all dark regions including the cavities between fingers are converted into white pixels, and all hand regions are converted into black pixels. This image is processed morphologically to remove very small connected regions and also any holes in the white connected regions as shown in Fig.2.

Fig. 1: A typical Palmprint image

Fig. 2: Cropped section of Fig.1

2. ROI Extraction
2.1 Creation of Palmprint Database This database was created at the Biometrics Research Laboratory at the Indian Institute of Technology Delhi. A database of palmprint images is collected by using a setup with Canon A630 digital camera. The setup includes a ring shaped fluorescent lighting source. The images are acquired online using a PC which is connected with the digital camera. Palmprint images of more than 100 users were collected. Users are staff/students of IIT Delhi. You can find palmprints of users of varying ages from 17 to 60 years in the database. A minimum of 7 images per user per hand were captured and the angle of rotation of the hand was changed for each capture. Some of the ideas for the

Fig. 3: Binarised Image .

Fig. 4: A common tangent between two curves

6.

On the left edge of the binary image, cavities are analyzed. Based on the rotation of the hand the two cavities one between the little finger and the ring finger, and another between the middle finger and the index finger are detected.

400

7. 8.

An edge detection filter (Laplacian) is applied on the fingers to get the curves of the cavities. A common tangent is formed between the two curves of the cavities in Fig. 4. A common tangent is drawn to the two curves such that all points of both the curves lie on one side of the tangent. The perpendicular bisector of this tangent is taken as x-axis and the tangent is taken as the y-axis. By defining this coordinate system one is able to extract the ROI at a fixed distance.

The results of authentication superimposed detail images.

correspond

to

9.

10. The ROI is rotated and cut to the relevant size and saved as a file.

Fig. 6: Two dimensional one level DWT decomposition

Fig. 5: The extracted ROI of palmprint

As we have five decompositions there will be five superimposed detail images of different sizes (i.e. the size of 1st detail image is 86x86, the 2nd detail image is 43x43, 3rd detail image 22x22, 4th detail image is 11x11 and 5th detail image is 6x6). These detail images are divided into non overlapping windows [1]. From each window, energy feature is calculated as:

3. Wavelet Features
Wavelets are immensely suitable for texture analysis. As a palmprint is a rich source of texture, we are inclined to investigate the effectiveness of Haar wavelets on the palmprints. Simplicity of these wavelets is an overriding factor in the choice of these wavelets though other celebrated wavelets such as Daubechies wavelets, Symlets are also tried but the improvement accrued from the latter ones is not commensurate with the effort. A two-dimensional discrete wavelet transform is applied on the ROI of a palmprint resulting in four subbands: HL1, LH1 and HH1, which correspond to fine-level coefficients and LL1, which corresponds to coarse-level coefficients. Note that HL1, LH1 and HH1 together contribute to a detail image. In this work, this detail image is used for representing the high frequency structures, i.e. lines, creases, wrinkles, ridges on a palmprint. Fig. 6 shows one level of decomposition. LL1 can be further decomposed into four subbands and the process is repeated till the information from LLn at the nth decomposition is insignificant. In this work, a palprint image is decomposed into five time scales using Haar wavelets beyond which results are found to be deteriorating. From these decompositions, detail images are obtained. In any decomposition, the detail images include horizontal, vertical and diagonal parts of the decomposed image. These three detail images are superimposed to yield one. It may be noted that as the decomposition level increases the size of the detail images also decreases.

Si ( x, y ) = H i ( x, y ) + Vi ( x, y ) + Di ( x, y )
Eid = ( Si ( x, y )) 2 , i = 1, 2,..., 5
x =1 y =1 M N

(1) (2)

where i is the decomposition level.

H i , Vi and Di are

the detail images in horizontal, vertical and diagonal directions respectively. Other features are under investigation.

Fig. 7: Three-level DWT decomposition

4. Fuzzy Features
The palmprints possess random textures consisting of zig zag and crisscrossing lines in addition to three principal lines. We are mainly interested in capturing the random texture. One way is to partition the ROI of palmprint into windows of appropriate size and represent the information by fuzzy features with respect to the centre of the window. The fuzzy features of the neighborhood are aggregated to yield a representative value called cumulative response

401

that represents the texture. The formulae used are given as part of the algorithm outlined below. 4.1 An Algorithm for Feature Extraction Fuzzy features are extracted in the following way: 1. 2. The ROI is divided into some number of nonoverlapping windows. Considering the center pixel I(i,j) in a window of size WxW, the average intensity as

between the same user and the imposter scores are derived from the matching distances between two different users. Given two data sets, a matching algorithm determines the degree of similarity between them. To describe the matching process clearly, we use a feature vector that consists of two feature matrices to represent the image data. A Euclidean distance is adopted to determine the similarity for palmprint matching. Let P and Q be two palmprint feature vectors. The Euclidean distance can P= ( p1, p2,...,pn) and Q = (q1 , q2 ,..., qn ) , be described as: If then the Euclidean distance from P to Q is:
d ( p , q ) = ( p1 q1 ) 2 + ( p2 q2 ) 2 + ... + ( pn qn ) 2

I (i,j) is defined
(3)

I (i , j ) =
3. 4.

WxW 1 m , n i , j

I ( m, n)

(6)

The maximum intensity, max I(i,j) in the window is found. Replacing the center pixel with I (i,j) and ij for every pixel is computed except the center pixel.

where n is the number of features.

5.1 Implementation Extraction of fuzzy and wavelet features is undertaken from ROI of palm images. We have taken 7 samples from each user and there are 125 users in all with a total of 125X7=875 images. Several experiments are conducted by taking different combinations of training and testing samples. The size of window is also varied to see its effect on the scores. The recognition scores corresponding to different features at different window sizes at different scales (or decompositions) are shown in Table1 where we have 4 samples for training and one sample for testing. Note that results of authentication using fuzzy feature is superior to those with wavelet feature. Table 2 gives recognition scores for different combinations of training and testing samples using only fuzzy feature. The matching score decrease as the number of testing samples is increased.
Table 1: The Recognition score (%) results with 4:1 Window size (Pixels) Fuzzy feature (%) Wavelet feature Scale 2(%) Scale 3(%) Scale 4(%) Scale 5(%)

I (i , j ) I (i , j ) ij = 1 max I (i, j )
5.

(4)

For every window, the cumulative response is computed as follows:

FF =

( xI (i, j ) )
ij

(5)

ij

The above value is used as a feature and number of these features is equal to the number of windows in a ROI.

5. Results
The matching algorithm performs a decision using some parameters (e.g. a threshold). The false acceptance rate (FAR) and the false rejection rate (FRR) can typically be traded off against each other by changing the threshold. We use genuine acceptance rate (GAR) to measure the performance of an authentication system, which depends on FAR. The graph of receiver operating characteristic (ROC) is a plot of GAR Vs. FAR obtained by changing the threshold value implicitly. FAR is defined as ratio of the number of successful fraud attempts to the total number of fraud attempts against a person and it measures the percent of invalid matches. FRR can be defined as the number of rejected verification attempts for a qualified person to the number of all verification attempts for a qualified person and it measures the percent of valid inputs being rejected. The number of successful fraud attempts and the rejected attempts for a qualified person are calculated by finding the genuine and imposter scores. Genuine scores are derived from the matching distances

11x11 9x9 7x7 5x5

99.2 100 100 100

41.6 68.0 80.8 88.0

52.8 75.2 87.2 93.6

52.0 75.2 86.4 94.4

94.4

Table 2: The Recognition score (%) with 4:1,5:2 &3:2 Window size (Pixels) Fuzzy feature (%) 4 train 1 test 5 train 2 test 3 train 2 test

11x11 9x9 7x7 5x5

99.2 100 100 100

95.6 95.6 96 96

82.4 87.2 87.2 88.4

We now show the plots of fuzzy features from all windows belonging to samples of two different users in

402

Fig. 8 and two samples belonging to the same user in Fig. 9 to get an idea of how they differ. The differences between the features of two different users and the same users are depicted in Fig. 10. It can be clearly seen that the differences are accentuated when the users are different as against the differences between the features of the same users. The effectiveness of the features can be judged from the relative differences. It is an easy matter to know that the relative differences from the wavelet feature will not be so large hence ROC due to wavelet feature points to the low matching scores as in Fig.12.
1 User1 User2

Fig. 10: The difference of Fuzzy feature vector in Figs. 8 and 9

0.9

0.8 Feature Vector

0.7

Fig. 11: ROC plot for three different combinations of training and testing samples using fuzzy feature with window size (5x5) pixels [3, 4].

0.6

0.5

0.4

50

100 150 Numbe of Feature

200

Fig. 8: Fuzzy Feature plot of two different users

1 Feature plot of image1 Feature plot of image2

0.9

0.8 Feature Vector

0.7

Fig. 12: ROC plot for wavelet feature at two-different decomposition level with window size (5x5) pixels.

0.6

0.5

0.4

50

100 150 Number of Feature

200

Fig. 9: Fuzzy Feature plot of the same user

0.25 Difference between the feature vector from same palm Difference between the feature vector from different palm 0.2

To compare the ROC characteristics arising out of fuzzy feature and wavelet feature due to different training and testing samples, plots are provided in Fig. 11 and Fig. 12 respectively. ROC characteristics from fuzzy feature are superior to those from wavelet feature. These plots reveal the fact that with more training samples and less testing samples, the matching scores are improved. This confirms the truth with more number of training samples the information gathered will be more and it will serve as a discriminating power of the authentication system.

6. Conclusions
A new palmprint acquisition is developed in Biometrics lab of IIT Delhi. A remarkable feature of this system is that the central portion of the palm is uncovered whereas the rest of the palmprint and fingers rest on the glass. The ROI is extracted from the uncovered portion of the palmprint. The ROI is crucial phase of the palmprint authentication system as it must be invariant to size, rotation and translation. The proposed cropping of the ROI from the acquired image meets the requirements of invariance. There are occasions when

Difference

0.15

0.1

0.05

50

100 150 Number of Feature

200

403

shadows hinder with the determination of tangent thus leading to wrong ROI. This case is very rare so no special provision is made to tackle this kind of problems. The effectiveness of features depends on the window size used to partition the ROI. After experimentation a window size of 9x9 pixels is found to give the best matching score using fuzzy feature. It is remarkable that one novel fuzzy feature devised by us is sufficient to represent the entire information of the palmprint ROI so well that it has discouraged us from exploring other features. We have also come up with an innovative idea in the extraction of a wavelet feature. Though multi-time scale decomposition has been extensively used in the separation of approximate and detail images from a given images, all the detail images have not been used in one go. In the proposed scheme, all the detail images arising out of decomposition of an image are superimposed as this superimposed image contains the high frequency components such as creases, principal lines, ridges etc. This image is partitioned into nonoverlapping windows for extracting an energy feature. There is an innumerable number of texture features in the context of texture analysis but these are not explored in this work. The future work will try out some additional features along with energy feature. In the wavelet case too the window size matters a lot. A window size of 5x5 is found to be appropriate out of a few sizes experimented. The ROC characteristics of palmprint based authentication system demonstrate the effectiveness of fuzzy feature over wavelet feature. Presently, we are experimenting additional features to consolidate our gains from the proposed features in this paper. The main problem with authentication of palmprint is the choice number of samples for training and testing. We have made several experiments by varying these numbers. It is observed that as the training samples increase the matching scores increase but as the number of testing samples increase the matching scores decrease correspondingly.

[4]

Anil K. Jain,Arun Ross and Salil Prabhakar,An Introduction to Biometric Recognition, IEEE Trans. on Circuits and System for Video Technology, Special Issue on Image and VideoBased Biometrics , Vol. 14, No. 1,January 2004.

N. Duta, A.K. Jain, and K.V. Mardia, Matching of Palmprint, Pattern Recognition Letters, Vol. 23, No. 4, pp. 477485, 2001. [6] David Zhang, Wai-Kin Kong, Jane You, and Michael Wong, Online Palmprint Identification, IEEE Trans. Pattern Analysis & Machine Intelligence, Vol. 25, No. 9, pp. 1041 1050, September 2003. [7] Ying-Han Pang, Andrew Teoh Beng Jin, David Ngo Chek Ling, "Palmprint based Cancelable Biometric Authentication System", Int. J. of Signal Processing, Vol.1, No.2, pp.93-99, 2004. [8] Ying-Han Pang, Andrew Teoh Beng Jin, David Ngo Chek Ling, Fu San Hiew, "Palmprint Verification with Moments", International Conference, 2004. [9] Lei Zhang, Zhenhua Guo, Zhou Wang, David Zhang, "Palmprint Verification Using Complex Wavelet Transform", IEEE In.. Conf. on Image Processing, San Antonio, TX, Sept.16-19,2007. [10] Li Fang, Maylor K.H. Leung , Tejas Shikhare, Victor Chan, Kean Fatt Choon , Palmprint Classification, IEEE Int. Conf. on Systems, Man and Cybernetics, Vol. 4, 8-11 October 2006, pp. 2965-2969.

[5]

References
[1] Xiang-Qian Wu, Kuan-Quan Wang, David Zhang, Wavelet Based Palmprint Recognition, First international conference on Machine learning and cybernetics, Beijing, pp-1253-1257, 4-5 Nov 2002 R. C. Gonzalez, R. E. Woods, Digital Image Processing, Addison Wesley publishers,1993 L. Hong and A. K. Jain, Integrating Faces and Fingerprints for Personal Identification, IEEE Trans. on Pattern Analysis and Machine Intelligence, Vol. 20, No. 12, pp. 1295-1307,Dec 1998.

[2] [3]

404