Palmprnt Feature Extracton Usng 2-D Gabor Flters Wa Kn Kong Davd Zhang and Wenxn L Bometrcs Research Centre Department of Computng The Hong Kong Polytechnc Unversty Kowloon Hong Kong Correspondng author: Prof. Davd Zhang Bometrcs Research Centre Department of Computng The Hong Kong Polytechnc Unversty Hung Hom Kowloon Hong Kong Phone: (852 2766-7271 Fax: (852 2774-0842 E-mal: csdzhang@comp.polyu.edu.hk
Abstract Bometrc dentfcaton s an emergng technology that can solve securty problems n our networked socety. A few years ago a new branch of bometrc technology palmprnt authentcaton was proposed [1] whereby lnes and ponts are extracted from palms for personal dentfcaton. In ths paper we consder the palmprnt as a pece of texture and apply texture-based feature extracton technques to palmprnt authentcaton. A 2-D Gabor flter s used to obtan texture nformaton and two palmprnt mages are compared n terms of ther hammng dstance. The expermental results llustrate the effectveness of our method. Index Terms Palmprnt Gabor flter bometrcs feature extracton texture analyss 1. Introducton Computer-based personal dentfcaton also known as bometrc computng whch attempts to recognze a person by hs/her body or behavoral characterstcs has more than 30 years of hstory. The frst commercal system called Identmat whch measured the shape of the hand and the length of fngers was developed n the 1970s. At the same tme fngerprnt-based automatc checkng systems were wdely used n law enforcement. Because of the rapd development of hardware ncludng computaton speed and capture devces rs retna face voce sgnature and DNA have oned the bometrc famly [2-324]. Fngerprnt dentfcaton has drawn consderable attenton over the last 25 years. However some people do not have clear fngerprnts because of ther physcal work or problematc skn. Irs and retna recognton provde very hgh accuracy but suffer from hgh costs of nput devces or ntruson nto users. Recently many researchers have focused on face and voce verfcaton systems; nevertheless ther performance s stll far from satsfactory [20]. The accuracy and unqueness of 3-D hand geometry are stll open questons 1
[2420]. Compared wth the other physcal characterstcs palmprnt authentcaton has several advantages: 1 low-resoluton magng; 2 low ntrusveness; 3 stable lne features and 4 hgh user acceptance. Palmprnt authentcaton can be dvded nto two categores on-lne and off-lne. Fg. 1 (a and (b show an on-lne palmprnt mage and an off-lne palmprnt mage respectvely. Research on off-lne palmprnt authentcaton has been the man focus n the past few years [1522-23] where all palmprnt samples are nked on paper then transmtted nto a computer through a dgtal scanner. Due to the relatve hgh-resoluton off-lne palmprnt mages (up to 500 dp some technques appled to fngerprnt mages could be useful for off-lne palmprnt authentcaton where lnes datum ponts and sngular ponts can be extracted [15]. For onlne palmprnt authentcaton the samples are drectly obtaned by a palmprnt scanner [25-26]. Recently a CCD based palmprnt capture devce has been developed by us [25]. Fg. 2(a shows a palmprnt mage captured by our palmprnt scanner and Fg. 2(b shows the outlook of the devce. Note that a low-resoluton technque (75 dp s adopted to reduce the mage sze whch s comparable wth fngerprnt mages even though a palm s much larger than a fngerprnt. It s evdent that on-lne dentfcaton s more mportant for many real-tme applcatons so that t draws our attenton to nvestgate. Our on-lne palmprnt verfcaton system contans fve modules palmprnt acquston preprocessng feature extracton matchng and storage. Fg. 3 gves a block dagram to descrbe the relatonshp between the fve modules. The fve modules are descrbed below: 1 Palmprnt Acquston: A palmrpnt mage s captured by our palmprnt scanner and then the AC sgnal s converted nto a dgtal sgnal whch s transmtted to a computer for further processng. 2
2 Preprocessng: A coordnate system s set up on bass of the boundares of fngers so as to extract a central part of a palmprnt for feature extracton. 3 Textured Feature Extracton: We apply a 2-D Gabor flter to extract textural nformaton from the central part. 4 Matchng: A dstance measure s used to measure the smlarty of two palmprnts. 5 Database: It s used to store the templates obtaned from the enrollment phase. Our palmprnt authentcaton system can operate n two modes enrollment and verfcaton. In the enrollment mode a user s to provde several palmprnt samples to the system. The samples are captured by our palmprnt scanner and passes through preprocessng and feature extracton to produce the templates stored n a gven database. In the verfcaton mode the user s asked to provde hs/her user ID and hs/her palmprnt sample. Then the palmprnt sample passes through preprocessng and feature extracton. The extracted features are compared wth templates n the database belongng to the same user ID. In ths paper we attempt to apply a textural extracton method to palmprnt mages for personal authentcaton. The remanng sectons are organzed as follows: preprocessng steps are mentoned n Secton 2. Palmprnt feature extracton by texture analyss s explaned n Secton 3. Expermental results are gven n Secton 4. Fnally Secton 5 summares the man results of ths paper. 3
2. Palmprnt Image Preprocessng Before feature extracton t s necessary to obtan a sub-mage from the captured palmprnt mage and to elmnate the varatons caused by rotaton and translaton. The fve man steps of palmprnt mage preprocessng are as follows (see Fg. 4. Step 1: Apply a low-pass flter to the orgnal mage. Then use a threshold T p to convert ths orgnal mage nto a bnary mage as shown n Fg. 4(b. Mathematcally ths transformaton can be represented as B(x y=1 f B(x y=0 f O( x y * L( x y T (1 p O ( x y * L( x y < T (2 p where B(xy and O(xy are the bnary mage and the orgnal mage respectvely; L(xy s a lowpass flter such as Gaussan and represents an operator of convoluton. Step 2: Extract the boundares of the holes (F x F y (=12 between fngers usng a boundary-trackng algorthm. The start ponts (Sx Sy and end ponts (Ex Ey of the holes are then marked n the process (see Fg. 4(c. Step 3: Compute the center of gravty (Cx Cy of each hole wth the followng equatons: Cx M ( = = 1 F x M( (3 Cy M ( = = 1 F y M( (4 where M( represents the number of boundary ponts n the hole. Then construct a lne that passes through (Cx Cy and the mdpont of (Sx Sy and (Ex Ey. The lne equaton s defned as 4
( Cy y = x ( Cx My MyCx Mx Cx Mx Cy Mx (5 where (Mx My s the mdpont of (Sx Sy and (Ex Ey. Based on these lnes two key ponts (k 1 k 2 can easly be detected (see Fg. 4(d. Step 4: Lne up k 1 and k 2 to get the Y-axs of the palmprnt coordnate system and make a lne through ther md pont whch s perpendcular to the Y-axs to determne the orgn of the coordnate system (see Fg. 4(e. Ths coordnate system can algn dfferent palmprnt mages. Step 5: Extract a sub-mage wth the fxed sze on the bass of coordnate system whch s located at the certan part of the palmprnt for feature extracton (see Fg. 4(f. 3. Palmprnt Feature Extracton By Texture Analyss Ths secton defnes our palmprnt feature extracton method whch ncludes flterng and matchng. The motvaton for usng a Gabor flter n our palmprnt research s frst dscussed. 3.1 Gabor Functon Gabor flter Gabor flter bank Gabor transform and Gabor wavelet are wdely appled to mage processng computer vson and pattern recognton. Ths functon can provde accurate tme-frequency locaton governed by the Uncertanty Prncple [6-7]. A crcular 2- D Gabor flter n the spatal doman has the followng general form [8-9] 1 2 2 x y G ( x y θ u σ = exp exp{ 2π( ux cosθ uy snθ } 2 2 (6 2πσ 2σ where = 1 ; u s the frequency of the snusodal wave; θ controls the orentaton of the functon and σ s the standard devaton of the Gaussan envelope. Such Gabor flters have 5
been wdely used n varous applcatons [10-19]. In addton to accurate tme-frequency locaton they also provde robustness aganst varyng brghtness and contrast of mages. Furthermore the flters can model the receptve felds of a smple cell n the prmary vsual cortex. Based on these propertes n ths paper we try to apply a Gabor flter to palmprnt authentcaton. 3.2 Flterng and Feature Extracton Generally prncpal lnes and wrnkles can be observed from our captured palmprnt mages (see Fg. 1(a. Some algorthms such as the stack flter [21] can obtan the prncpal lnes. However these lnes do not contrbute adequately to hgh accuracy because of ther smlarty amongst dfferent palms. Fg. 5 shows sx palmprnt mages wth smlar prncpal lnes. Thus wrnkles play an mportant role n palmprnt authentcaton but accurately extractng them s stll a dffcult task. Ths motvates us to apply texture analyss to palmprnt authentcaton. In fact a Gabor functon G ( x y θ u σ wth a specal set of parameters (σ θ u s transformed nto a dscrete Gabor flter G [ x y θ u σ ]. The parameters are chosen from 12 sets of parameters lsted n Table 1 based on the expermental results n the next secton. In order to provde more robustness to brghtness the Gabor flter s turned to zero DC (drect current wth the applcaton of the followng formula: n ~ G[ θ u σ ] = n = n G[ x y θ u σ ] = G[ x y θ u σ ] (7 2 (2n 1 n where (2n1 2 s the sze of the flter. In fact the magnary part of the Gabor flter automatcally has zero DC because of odd symmetry. Ths adusted Gabor flter wll convolute wth a sub-mage defned n Secton 2. The sample pont n the fltered mage s coded to two bts (b r b by the followng nequaltes 6
7 b r =1 f ] ]* [ ~ Re[ I u y x G σ θ 0 (8 b r =0 f < ] ]* [ ~ Re[ I u y x G σ θ 0 (9 b =1 f ] ]* [ ~ Im[ I u y x G σ θ 0 (10 b =0 f < ] ]* [ ~ Im[ I u y x G σ θ 0 (11 where I s the sub-mage of a palmprnt. Usng ths codng method only the phase nformaton n palmprnt mages s stored n the feature vector. The sze of the feature s 256 bytes. Fg. 6 shows the features generated by the 12 flters lsted n Table 1. Ths texture feature extracton method has been appled to rs recognton [13]. 3.3 Palmprnt Matchng In order to descrbe clearly the matchng process each feature vector s consdered as two 2- D feature matrces real and magnary. Palmprnt matchng s based on a normalzed hammng dstance. Let P and Q be two palmprnt feature matrces. The normalzed hammng dstance can be defned as ( 2 1 1 2 ( ( ( ( N Q P Q P D N N I I R R o = = = (12 where P R (Q R and P I (Q I are the real part and the magnary part of P (Q respectvely; the Boolean operator s equal to zero f and only f the two bts P R(I ( and Q R(I ( are equal and the sze of the feature matrces s N N. It s noted that D o s between 1 and 0. The hammng dstance for perfect matchng s zero. In order to provde translaton nvarance matchng Eq. (12 can be mproved as ( ( ( 2 ( ( ( ( mn mn( max(11 mn( max(11 mn t H s H Q t s P Q t s P D s N N s t N N t I I R R T t S s = = = < < (13
where S=2 and T=2 control the range of horzontal and vertcal translaton of a feature n the matchng process respectvely and H ( s = mn( N N s max(1 1 s. (14 The hammng dstance D mn can support translaton matchng; nevertheless because of unstable preprocessng t s not a rotatonal nvarant matchng. Therefore n enrollment mode the coordnate system s rotated by a few degrees and then the sub-mages are extracted for feature extracton. Fnally combnng the effect of preprocessng and rotated features Eq. (13 can provde both approxmately rotatonal and translaton nvarance matchng. 4. Expermental Results 4.1 Palmprnt Database In the followng experments a palmprnt database contans 4647 palmprnt mages collected from 120 ndvduals by usng our palmprnt scanner. 43 of them are female 111 of them are less than 30 years old and 2 of them are more than 50 years old. Each of them s asked to provde about 10 mages for ther left palm and 10 mages for ther rght palm n each of two occasons. In total each subect provdes about 40 mages. The average tme dfference of the frst and second occasons s 81 days. The maxmum and mnmum are 4 and 162 days respectvely. Orgnally the collected mages have two szes 384 284 and 768 568. The large mages are reszed to 384 284; consequently the sze of all the test mages n the followng experments s 384 284 wth 75dp resoluton. The central parts of each mage extracted wth sze 128 by 128 are named DBI. The preprocessed mages n DBI reszed to 64 by 64 are named DBII. The DBII s used to test the possblty of usng lower-resoluton palmprnt mages for personal dentfcaton. Fg. 7 shows nne typcal mages from our databases. 8
4.2 Verfcaton Tests To obtan better parameters for our system 12 dfferent sets of parameters lsted n Table 1 are used to test the method. The frst four flters are named as Level 1 flters snce they are dfferent n the parameter θ. Smlarly the other flters are named as Level 2 and Level 3. Each of the mages n DBI (DBII s matched wth all other palmprnt mages n the same database. A matchng s counted as a correct matchng f two palmprnt mages are collected from the same palm; otherwse t s an ncorrect matchng. The total number of matchngs for one verfcaton test s 10794981. 43660 of them are correct matchngs and rest of them are ncorrect matchngs. In total 24 verfcaton tests are carred out for testng the 12 sets of parameters on the two databases. The performance of dfferent parameters on the two databases s presented by Recever Operatng Characterstc (ROC curves whch are a plot of genune acceptance rate aganst false acceptance rate for all possble operatng ponts. Fgs. 8(a 8(b and 8(c (8(d 8(e and 8(f show the ROC curves for DBI (DBII generated by Levels 1 2 3 flters respectvely. Accordng to the ROC curves Level 3 flters are better than Levels 1 and 2 flters for DBI. Accordng to Fg. 8(c Flters 9 10 and 11 provde smlar performance when the false acceptance rate s greater than 10-2. For false acceptance rates smaller than 10-2 Flters 9 and 11 are better than Flter 10. For DBII Level 1 flters are better than Level 2 and 3 flters. In fact Flter 2 s the best for DBII. Although usng very low-resoluton mages as DBII s mages cannot provde very good performance t stll gves us an nsght nto usng very low-resoluton palmprnt mages for personal dentfcaton. 5. Conclusons Ths paper reports a textured-based feature extracton method usng low-resoluton palmprnt mages for personal authentcaton. A palmprnt s consdered as a texture mage so an 9
adusted Gabor flter s employed to capture the texture nformaton on palmprnts. Based on our tests Flter 11 s the best of twelve flters n terms of accuracy. Combned wth the effects of preprocessng and rotated preprocessed mages our matchng process s translaton and rotatonal nvarance. Expermental results llustrate the effectveness of the method. Acknowledgments The work s partally supported by the UGC (CRC fund from Hong Kong Government and the central fund from The Hong Kong Polytechnc Unversty. References [1] D. Zhang and W. Shu Two novel characterstcs n palmprnt verfcaton: datum pont nvarance and lne feature matchng Pattern Recognton vol. 32 no. 4 pp. 691-702 1999. [2] A. Jan R. Bolle and S. Pankant (eds. Bometrcs: Personal Identfcaton n Networked Socety Kluwer Academc Publshers 1999. [3] D. Zhang Automated Bometrcs Technologes and Systems Kluwer Academc Publshers 2000. [4] R. Sanchez C. Sanchez-Avla and A. Gonzalez-Marcos Bometrc dentfcaton through geometry measurements IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 22 no. 10 pp. 1168-1171 2000. [5] W. Shu and D. Zhang Automated personal dentfcaton by palmprnt Optcal Engneerng vol. 37 no. 8 pp.2659-2362 1998. [6] D. Gabor Theory of communcatons J. IEE vol. 93 pp. 429-457 1946. [7] C.K. Chu An Introducton to Wavelets Academc Press Boston 1992 [8] J.G. Daugman Two-dmensonal spectral analyss of cortcal receptve feld profles Vson Research vol. 20 pp. 847-856 1980. [9] J. Daugman Uncertanty relaton for resoluton n space spatal frequency and orentaton optmzed by two-dmensonal vsual cortcal flters Journal of the Optcal Socety of Amerca A vol. 2 pp. 1160-1169 1985. 10
[10] A. Jan and G. Healey A multscale representaton ncludng opponent color features for texture recognton IEEE Transactons on Image Processng vol. 7 no. 1 pp. 124-128 1998. [11] D. Dunn and W.E. Hggns Optmal Gabor flters for texture segmentaton IEEE Transactons on Image Processng vol. 4 no. 4 pp. 947-964 1995. [12] A.C. Bovk M. Clark and W.S. Gesler Multchannel texture analyss usng localzed spatal flters IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 12 no. 1 pp. 55-73 1990. [13] J. Daugman Hgh confdence vsual recognton of persons by a test of statstcal ndependence IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 15 no. 11 pp. 1148-1161 1993. [14] A.K. Jan S. Prabhakar L. Hong and S. Pankant Flterbank-based fngerprnt matchng IEEE Transactons on Image Processng vol. 9 no. 5 pp. 846-859 2000. [15] L. Hong Y. Wan and A. Jan Fngerprnt mage enhancement algorthm and performance evaluaton IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 20 no. 8 pp. 777-789 1998. [16] C.J. Lee and S.D. Wang. Fngerprnt feature extracton usng Gabor flters Electronc Letters vol. 35 no. 4 pp. 288-290 1999. [17] M.J. Lyons J. Budynek and S. Akamatsu Automatc classfcaton of sngle facal mages IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 21 no. 12 pp. 1357-1362 1999. [18] B. Duc S. Fscher and J. Bgun Face authentcaton wth Gabor nformaton on deformable graphs IEEE Transactons on Image Processng vol. 8 no. 4 pp. 504-516 1999. [19] Y. Adn Y. Moses and S. Ullman Face recognton: The problem of compensaton for changes n llumnaton drecton IEEE Transactons on Pattern Analyss and Machne Intellgence vol. 19 no. 7 pp. 721-732 1997. [20] S. Pankant R.M. Bolle and A. Jan Bometrcs: The Future of Identfcaton IEEE Computer vol. 33 no. 2 pp. 46-49 2000. [21] P.S. Wu and M. L. Pyramd edge detecton based on stack flter Pattern Recognton Letters vol. 18 no. 4 pp. 239-248 1997. [22] J. You W. L and D. Zhang Herarchcal palmprnt dentfcaton va multple feature extracton Pattern Recognton vol. 35 no. 4 pp. 847-859 2002. [23] N. Duta A.K. Jan and K.V. Marda Matchng of Palmprnt Pattern Recognton Letters vol. 23 no. 4 pp. 477-485 2001. 11
[24] D. Zhang (ed. Bometrc Resolutons for Authentcaton n an e-world Kluwer Academc Publshers 2002. [25] D. Zhang W.K. Kong J. You and M. Wong On-lne palmprnt dentfcaton To be appear n IEEE Transactons on Pattern Recognton and Machne Intellgence. [26] C.C. Han H.L. Cheng K.C. Fan and C.L. Ln Personal authentcaton usng palmprnt features Pattern Recognton Specal Issue: Bometrcs vol. 36 no 2 pp. 371-381 2003. 12
Fgures: Fg. 1 Fg. 2 Fg. 3 Fg. 4 Fg. 5 Examples of (a on-lne wth lne defntons and (b off-lne palmprnt mages. Capture devce and captured palmprnt mages. (a On-lne palmprnt mage obtaned by our palmprnt scanner and (b our palmprnt capture devce. Block dagram of our palmprnt verfcaton system. Man steps of preprocessng. (a Orgnal mage (b Bnary mage (c Boundary trackng (d Key ponts (k 1 and k 3 detectng (e The coordnate system and (f The central part of a palmprnt. Three sx mages wth smlar prncpal lnes Fg. 6 Orgnal mage from DBI and ther features generated by 12 flters lsted n Table. a Orgnal mage b d and f real parts of features from Level 1 2 and 3 flters respectvely c e and g magnary parts of from Level 1 2 and 3 flters respectvely. Fg. 7. Nne typcal mages from DBI. Fg. 8 Verfcaton test results. (a (b and (c ((d (e and (f the ROC curves of Level 1 Level 2 and Level 3 flters from DBI (DBII respectvely Tables: Table 1 The parameters of the 12 flters. 13
Fgures: (a Fg. 1 (b 14
(a (b Fg. 2 15
Fg. 3 16
(a (d (b (e (c (f Fg.4 17
Fg. 5 18
(a (b (c (d (e (f (g Fg. 6 19
Fg. 7 20
(a (b (c (d (e (f Fg. 8 21
Tables Table 1 Levels No Szes θ u σ 1 9 by 9 0 0.3666 1.4045 1 2 9 by 9 45 0.3666 1.4045 3 9 by 9 90 0.3666 1.4045 4 9 by 9 135 0.3666 1.4045 5 17 by 17 0 0.1833 2.8090 2 6 17 by 17 45 0.1833 2.8090 7 17 by 17 90 0.1833 2.8090 8 17 by 17 135 0.1833 2.8090 9 35 by 35 0 0.0916 5.6179 3 10 35 by 35 45 0.0916 5.6179 11 35 by 35 90 0.0916 5.6179 12 35 by 35 135 0.0916 5.6179 22