International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India
Finger-Knuckle-Print Verification using Kekre’s Wavelet Transform H B Kekre
V A Bharadi
Computer Science Department MPSTME, NMIMS University Mumbai, India +91-9323557897
Computer Science Department MPSTME, NMIMS University Mumbai, India +91-9819125676
[email protected]
[email protected]
ABSTRACT
The biometrics is mainly concerned with ‘what you are’ rather than ‘what you carry’. The driving force of the progress in this field is, above all, the growing role of the Internet and the requirements of society [4],[5]. Therefore, considerable implementations can be found in the area of electronic commerce and electronic banking systems and security applications of vital installations [6].
Finger-Knuckle-Print is an emerging biometric trait. Because of high degree of uniqueness and low requirement of user cooperation it has got potential to become building block of future biometric security systems. In this paper we propose fingerknuckle-print verification using kekre’s wavelets. Kekre’s wavelets are based on kekre’s transform and easy to construct. We discuss the feasibility of using kekre’s wavelet to extract spectral features of the finger-knuckle-print and use them for verification. In this paper we have used wavelet energy feature, we have compared the results with Haar wavelets. Kekre’s wavelet based features give moderate accuracy and performance is same as Haar wavelets.
The biometrics has a significant advantage over traditional authentication techniques (namely smartcards, PIN numbers, passwords etc.) due to the fact that biometric characteristics of a person are not easily transferable, are unique of every person, and cannot be lost, stolen or broken. The choice of one of the biometric solutions depends on several factors [2]: User acceptance Level of security required Accuracy Cost and implementation time Fingerprint, Palm-print, Speech, Face, Iris, Retina, Hand Geometry, Ear Geometry are commonly used biometric means of human authentication [1],[2],[3],[4] besides there are other biometric traits also. Any part of human body is unique and we can use it as biometric trait provided we have a scanner for capturing that area.
Categories and Subject Descriptors I.4.7 Image Processing and Computer vision
General Terms
Measurement, Performance, Design, Experimentation, Security, Human Factors, Verification.
Keywords
Biometrics, Finger-Knuckle-Print, Kekre’s Wavelets.
Finger-knuckle-print (hereafter referred as FKP) is one of emerging biometric traits, as scanner or capturing hardware for this has been developed and database for research purpose is available [7]. The finger-knuckle-print (FKP) refers to the image of the outer surface of the finger phalangeal joint. Typical Knuckle-Print image is shown in Fig. 1, taken from the PolyU FKP database. Researchers at Hong Kong PolyU University have developed scanner for FKP [9] this is shown in Fig. 1 (a).
1. INTRODUCTION
With the advancement in the technology various methods of person identification are possible because of availability of affordable and co-operative sensors as well as capable processing computers and storage media. Another driving force behind widespread use of biometrics is the terrorism and need of better mechanism to protect sensitive areas where public interaction is more, like airports and railway stations. The methods are numerous, and are based on different personal characteristics. Voice, lip movements, hand geometry, face, odor, gait, iris, retina, fingerprint are the most commonly used authentication methods. All of these and behavioral characteristics are called biometrics [1], [2].
To implement a good biometric security system, we need to verify the biometric trait using feature vector [1],[2],[7],[8]. The feature vector is the information extracted from captured data. To extract the feature vector we need to segment the region of interest which contains information relevant for feature extraction, this is called as Region of Interest (ROI) [9],[10],[11]. A typical ROI extracted from FKP image is shown in Fig. 2.
The biometrics is most commonly defined as measurable psychological or behavioral characteristic of the individual that can be used in personal identification and verification [3], [4]. Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. ICWET’11, February 25–26, 2011, Mumbai, Maharashtra, India. Copyright © 2011 ACM 978-1-4503-0449-8/11/02…$10.00
(a) 32
International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India back surface is captured[8],[9],[10],[11]. Band limited phase only correlation function is proposed in [11] by them which give EER in the range of 5.5% to 0.31 %. In [9] they have proposed a localglobal feature fusion for FKP verification; Local features are extracted using a bank of Gabor filters convolved with FKP ROI and global features are taken from band limited phase only correlation function. In this paper propose use of wavelet based features, specifically wavelet energy of the FKP ROI for verification purpose. We are using a new family of wavelets called kekre’s wavelets. We discuss this in the next section.
(b)
3. KEKRE’S WAVELETS
Figure 1. (a) Finger Knuckle Print Capture Device [9] (b) Typical Finger-Knuckle-Print Image from Hong Kong Polytechnic University FKP Database [7].
Kekre’s wavelets are orthogonal family of wavelets. For generation of kekre’s wavelets we need basis function as in case of other families, this basis functions are generated from Kekre’s Transform matrix.
Wavelet Transform has been widely used to extract localized spectral information in the given data. In this paper we have used a new family of wavelets called kekre’s wavelets. The kekre’s wavelets are derived from kekre’s transform [12]. This will be discussed in coming sections. In the next section we discuss the different approaches for FKP verification.
3.1 Kekre’s Transform [12],[19]
Let us generate the Kekre’s Matrix [K] for size m x m where m can be any integer not necessarily the power of 2 as required for many other conventional transforms. This matrix has all 1’s on the main diagonal and upper triangle of the matrix. The sub-diagonal just below the main diagonal has the value (-m + i) where ‘m’ is the order of matrix and ‘i’ is the column number. Rests of the elements of lower triangle below the sub diagonal are all zeros. The general form of Kekre’s matrix [K] can be written as
Figure 2. Region of Interest (ROI) segmented from FKP shown in Fig. 1.
2. FINGER-KNUCKLE-PRINT BIOMETRIC TRAIT
AS
A
Among different kinds of biometric traits, hand-based biometrics has been attracting considerable attention over recent years. Fingerprint, palmprint, hand geometry, hand vein [1],[2],[4], and inner-knuckle-print [17],[18] have been proposed and discussed in the literature. The popularity & widespread use of hand-based biometrics should be attributed to its high user acceptance. In fact, the image pattern in the finger knuckle surface is highly unique and thus can serve as a distinctive biometric identifier [8]. FKP being recent has been yet to be thoroughly discovered. The current research has shown great potential in FKP to be used as an efficient and accurate biometric trait [8],[9],[10],[11].
The formula for generating the element Kxy of Kekre’s transform matrix is,
(1) The properties of Kekre’s Transform are as follows: 1) The Kekre’s transform is real and orthogonal transform. a. [K]T [ K]=[µ] (2) Where [k]T is transpose of [K] and [µ] is a diagonal matrix and its elements are given by µ11 = m µii = (m-i+1)(m-i+2) (3) 2) It has a fast algorithm as it contains m(m+1)/2 number of ones and (m-1)(m-2)/2 number of zeros leaving only (m-1) integer multiplications and only (m-1)(m/2) additions for transforming a column vector of dimension mx1. For a normal matrix transformation we require m2 multiplications and m (m-1) additions. 3) The transform of a vector f is given by F = [K] f (4) And inverse is given by
Hand geometry, especially 3D features from finger surface has been used in [13],[14] as a biometric traits but specific localized part has not been proposed. They used curvature based shape index to represent the finger back surface, rather than texture rich Finger-knuckle surface. Ravikanth & Kumar [15],[16] have proposed use of finger back surface as biometric feature; the whole back surface of hand is captured and then pre-processed to isolate the finger knuckle. They used subspace analysis using PCA & LDA for FKP analysis, in [10] L. Zang et al. have discussed this as a sub-optimal approach for FKP verification. L. Zang et al. have captured FKP using a specially developed device which essentially captures finger knuckle in a localized way, without much redundancy as only the concerned fingers 33
International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India f = [K]T [µ]-1 F
Values of matrix T can be computed as,
(5)
T(x, y) = K( N-P+(x+1), N-P+ y) ; 1≤ x≤ (P-1) , 1≤ y≤ P (3.43)
3.2 Kekre’s Wavelets [12]
Kekre’s Wavelet transform is derived from Kekre’s transform. From NxN Kekre’s transform matrix, we can generate Kekre’s Wavelet transform matrices of size (2N)x(2N), (3N)x(3N),……, up to maximum (N2)x(N2). For example, from 5x5 Kekre’s transform matrix, we can generate Kekre’s Wavelet transform matrices of size 10x10, 15x15, 20x20 and 25x25.
First row of T is used to generate (N+1) to 2N rows of Kekre’s Wavelet transform matrix. Second row of T is used to generate (2N+1) to 3N rows of Kekre’s Wavelet transform matrix, Like wise last row of T is used to generate ((P-1) N + 1) to PN rows [12]. We have used Kekre’s Wavelet Transform Matrices of Size 128, 64, 32 Generated from kekre’s Transform Matrix of Size 64, 32, 16 respectively. We calculate Wavelet energy feature for the Palmprint image using these wavelet matrices.
3.2.1 Properties of Kekre’s Wavelet Transform:
1. Orthogonal- The transform matrix K is said to be orthogonal if the following condition is satisfied. [K][K]T = [D],
Where D is a diagonal matrix.
Kekre’s Wavelet Transform matrix satisfies this property and hence it is orthogonal. The diagonal matrix value of Kekre’s transform matrix of size NxN can be computed as Figure 3. Kekre’s Transform (KT) matrix of size NxN. In general MxM Kekre’s Wavelet transform matrix can be generated from NxN Kekre’s transform matrix, such that M = N * P where P is any integer between 2 and N that is, 2 ≤ P ≤ N. Consider the Kekre’s transform matrix of size NxN shown in Fig. 3. MxM Kekre’s Wavelet transform matrix generated from NxN Kekre’s transform matrix is shown in Fig 4. First ‘N’ number of rows of Kekre’s Wavelet transform matrix is generated by repeating every column of Kekre’s transform matrix P times. To generate remaining (M-N) rows, extract last (P-1) rows and last P columns from Kekre’s transform matrix and store extracted elements in to temporary matrix say T of size (P-1) x P . Fig. 4 shows extracted elements of Kekre’s transform matrix stored in T.
(6) 2. Asymmetric- As the Kekre’s transform is upper triangular matrix, it is asymmetric. 3. Non Involutional - An involutionary function is a function that is its own inverse. So involution transform is a transform which is inverse transform of itself. Kekre’s transform is non-involution transform. 4. Transform on Vector -The Kekre’s Wavelet transform on a column vector f is given by
In general MxM Kekre’s Wavelet transform matrix can be generated from NxN Kekre’s transform matrix, such that M = N * P where P is any integer between 2 and N that is, 2 ≤ P ≤ N. Consider the Kekre’s transform matrix of size NxN shown in Fig. 3. MxM Kekre’s Wavelet transform matrix generated from NxN Kekre’s transform matrix is shown in Fig 4. First ‘N’ number of rows of Kekre’s Wavelet transform matrix is generated by repeating every column of Kekre’s transform matrix P times. To generate remaining (M-N) rows, extract last (P-1) rows and last P columns from Kekre’s transform matrix and store extracted elements in to temporary matrix say T of size (P-1) x P . Fig. 4 shows extracted elements of Kekre’s transform matrix stored in T.
F = [KW] f
(7)
And inverse is given by f = [KW]T [D]-1 F
(8)
5. Transform on 2D Matrix- Kekre’s Wavelet transform on 2D matrix f is given by [F] = [KW] [f] [KW]T
(9)
Obtaining Inverse: Calculate Diagonal matrix D as, [D] = [KW][KW]T
(10)
Inverse is calculated as [f] = [KW]T [ Fij / Dij ] [KW]
Figure 4.Temporary Matrix T of size (P-1) x P. 34
(11)
International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India Where Dij = Di * Dj
; 1≤ i ≤ N and 1≤ j ≤ N
we get total 9 Values. Such three blocks are there hence for a Finger-Knuckle-Print there are total 27 values in a feature vector. Currently we are using sum of all WE coefficients in a component (LH, HL or HH). This is shown in Fig. 6.
WE01 .. .. WE22 0 ,.., WKW .. (3.62) WE WE22 2 01
(a)
We have taken the values as it is without normalization. Such feature vectors are generated for both Kekre’s Wavelets and Haar wavelet. To find the matching between two FKP’s we take the Euclidian distance between two feature vectors W1 & W2 (both Kekre’s Wavelet & Haar).
(b) Figure 5. (a) Orientation of Bbocks for feature extraction (b) Three blocks of 128X128 Pixels.
ED
4. FKP FEATURE EXTRACTION USING KEKRE’S WAVELETS
n0
(Where N=27) (3.63)
5. FKP RECOGNITION RESULTS & DISCUSSIONS
Each block shown in Fig. 5(b) is of 128X128 Pixels size, now we can implement wavelet transform on it. We take wavelet transform up to 3 Levels (128, 64 and 32). At each level we have four set of wavelet coefficients LL, LH, LH, HH. LL component is used for next level after down sampling the inverse of LL only wavelet coefficients. We find the sum of coefficients in LH, HL & HH Blocks. This is given by [21],
We have used POLYU FKP [7] database for testing our method. The testing code is written in Microsoft Visual C# 2005 (.NET Framework 2.0). Testing is performed on AMD Athlon 64 FX 1.8 GHz processor, with 1.5GB RAM and Windows XP SP3 (32 Bit) Operating System. We have extracted feature vector using Kekre’s wavelet and Haar Wavelet. Total 512 tests were performed on FKP database. In this experiment we have mainly analyzed discrimination capacity of Kekre’s wavelets and their feasibility in spectral feature extraction.
W -1 W -1 i 0 j 0
1n W 2 n
The feature vector generation and distance measurement method discussed here will be used for FKP verification. We have extracted this feature vector for the segmented FKP ROI using Kekre’s Wavelets and Haar Wavelets, kekre’s wavelets performance is compared with that of Haar Wavelets. This is discussed in next section.
The Feature vector is extracted using Kekre’s Wavelets. We are using Wavelet Energy Feature [21]. We are using Kekre’s Wavelet and comparing the results with Haar Wavelets. The Region of Interest (ROI) is a texture rich surface (consisting of Wrinkles) as shown in Fig. 5(a). The wavelets will capture localized spectral information from the ROI. We have the ROI of Size 256 X 128 Pixels; we divide the ROI in three Blocks of 128 X 128 Pixels size as follows,
WELC WC[i, j ]2
N 1
W
(3.61)
We have performed intra-class and inter-class matching. This will be used to analyze genuine and imposter identification capability of the FKP based verification. The Kekre’s Wavelet Feature Vector’s distance for Genuine FKP matching (Intra-class) and imposter (Forgery) FKP matching (Inter-class) with the probability of occurrence is shown in Fig. 7. The graph shows that two different classes formed for the genuine and forgery class. The distance vs. probability plot clearly indicates that there are two separate classes of distance; the Genuine Distance class and Forgery (Imposter) Distance class, separated by a threshold of 475.
L=0, 1, 2; C = 0(LH), 1(HL), 2(HH) & W is the Size of Wavelet Component (64, 32, 16)
Furthermore we have performed TAR-TRR analysis [1],[2] for both Kekre’s Wavelets and Haar Wavelet based feature vector, we have performed this analysis on the data of above mentioned 512 tests. Euclidian distance based classification was used, it was found that Haar Wavelets and Kekre’s wavelets give same Equal Error Rate (EER)[1] of 80% for True Acceptance Rate (TAR) Vs. True Rejection Rate (TRR) analysis. The EER for False Acceptance Rate vs. False Rejection Rate (FRR) was found to be 20%; as shown in Fig. 8 (a) & (b).
Figure 6. Wavelet Coefficients blocks for Wavelet Energty Feature. We have three decomposition levels, and each level gives 3 Wavelet Energy (WE) values, hence for each block of FKP image 35
International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India
Figure 7. Kekre’s Wavelet Distance Probability Two separate classes belonging to Genuine and Forgery tests are clearly visible. Overlapping is higher due to localized coefficients are not consider but full componentwise sum is used for generating feature vector.
(a)
(b) 36
International Conference and Workshop on Emerging Trends in Technology (ICWET 2011) – TCET, Mumbai, India Figure 8. TAR Vs. TRR Plot for Kekre’s and Haar Wavelet [6]
Table I gives the summary of testing for Haar & Kekre’s Wavelets. Table 1 Performance Summary Type of EER (TAREER (FARSr. Wavelet TRR) FRR) Kekre’s 80% 20% 1 Wavelet Haar 80% 20% 2 Wavelet As compared to the systems suggested in [8],[9],[10],[11] the current system as lower performance level, but this is mainly due to the simpler classifier used. If a Neural Network based classifier with proper training is used the system can achieve better EER. Another thing is that we have used the Kekre’s Wavelet transform (KWT) for first time for pattern recognition, KWT is simpler to generate and involves only integer values in wavelet matrices hence faster calculations can be achieved; it gives results same as Haar Wavelets. This makes kekre’s Wavelet Transform very attractive for extraction for Spectral information for pattern Recognition.
Maltoni, D., Maio D., Jain A.K., Prabhakar S., “Handbook of Fingerprint Recognition”, Springer-Verlag, New York, 2003
[7] http://www.comp.polyu.edu.hk/~biometrics/FKP.htm [8] Lin Z.,Lei Z., D.Zhanga, H. Zhub, "Online Finger-KnucklePrint Verification for Personal Authentication", Biometrics Research Center, Department of Computing, The Hong Kong Polytechnic University [9] Lin Z., Lei Z., David Z., H. Zhub, "Ensemble of Local and Global Features for Finger-Knuckle-Print Recognition", Biometrics Research Center, Department of Computing, The Hong Kong Polytechnic University [10] Lin Zhang, Lei Zhang, D. Zhang, "Finger-Knuckle-Print: A New Biometric Identifier",Biometrics Research Center, Department of Computing, [11] Lin Zhang, Lei Zhang, D. Zhang, "Finger-Knuckle-Print Verification Based on Band-Limited Phase-Only Correlation",Biometrics Research Center, Department of Computing [12] H. B. Kekre, A. Athawale, D. Sadavarti, "Algorithm To Generate Kekre’s Wavelet Transform from Kekre’s Transform",IJSET, June 2010 (In Press)
6. CONCLUSION
In this paper we have proposed use of a new family of wavelets called as Kekre’s Wavelets for Finger-Knuckle-Print Verification. Kekre’s Wavelet are faster than Haar wavelet as the Transform Matrix has only integer values. The proposed system gives moderate EER of 80% bute performance is equal to Haar Wavelets. This indicates that kekre’s wavelets have great potential to be used for spectral feature extraction. The performance can be improved by using neural network based classifier with proper training.
[13] Woodard D., Flynn J.,"Personal identification utilizing finger surface features",Proc. CVPR, vol. 2, pp. 1030–1036 (2005). [14] Woodard D., Flynn J.,"Finger surface as a biometric identifier", Computer Vision and Image Understanding 100(3), 357–384, (2005) [15] Ravikanth C., Kumar A.,"Personal authentication using finger knuckle surface", IEEE Trans. Information Forensics and Security 4(1),98–109,(2009)
7. ACKNOWLEDGMENTS
[16] Ravikanth C., Kumar A.,"Biometric authentication using finger-back surface",Proceedings of CVPR, pp. 1–6 (2007)
Authors are very thankful to L. Zhang, Assistant Professor, Biometric Research Centre (UGC/CRC), The Hong Kong Polytechnic University, for providing the Finger-Knuckle-Print database. This database has been a key resource for this research.
[17] Q. Li, Z. Qiu, D. Sun, J. Wu, Personal identification using knuckle print, in: Proceedings of SinoBiometrics,2004, pp. 680-689
8. REFERENCES
[18] L. Nanni, A. Lumini, A multi-matcher system based on knuckle-based features, Neural Computing & Applications 18 (1) (2009) 87-91.
[1] A. K. Jain, A. Ross, S. Prabhakar, “An Introduction to Biometric Recognition”, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 1, January 2004
[19] H. B. Kekre, S. D. Thepade, “Image Retrieval using NonInvolutional Orthogonal Kekre’s Transfrom”, International Journal of MultiDisciplinary Research And Advnces in Engineering,IJMRAE, Vol.1, No.I,Novenber 2009,pp189203.
[2] A. K Jain, P. Flynn, A. Ross, “ Handbook of Biometrics”, Springer, ISBN-13: 978-0-387-71040-2 ,2008 [3] C. Tisse, L. Martin, L. Torres, and M. Robert, “Person Identification Technique using Human Iris Recognition”, in Proc. Vision Interface, pp. 294-299, 2002.
[20] H. B. Kekre, V A Bharadi, “Finger-Knuckle-Print Region of Interest Segmentation using Gradient Field Orientation & Coherence”, IEEE CNF, ICETET 2010, India (Unpublished)
[4] A. K. Jain, A. Ross, S. Prabhakar, “An Introduction to Biometric Recognition”, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 1, January 2004
[21] K Y Edward Wong,G. Sainarayanan, A. Chekima,"Palmprint Identification Using Wavelet Energy", International Conference on Intelligent and Advanced Systems 2007,pp 714-719, DOI: 1-4244-1355-9/07
[5] Arivazhagan S, Mumtaj J.,Ganesan L., "Face Recognition using Multi-Resolution Transform", International Conference on Computational Intelligence and Multimedia Applications 2007, IEEE DOI 10.1109/ICCIMA.2007
37
