Principal Component Analysis for Minutiae Verification on Fingerprint Image

Transcription

1 roceedngs of the 7th WSEAS Internatonal Conference on Multeda Systes & Sgnal rocessng, Hangzhou, Chna, Aprl 5-7, rncpal Coponent Analyss for Mnutae Verfcaton on Fngerprnt Iage CHING-ANG HSIEH, SHYS-RONG SHYU Departent of Electrcal Engneerng akang Unversty 5 Yng-chuan Road, asu, ape County, awan 537 AIWAN el: , Fax: Abstract: Mnutae are the two ost pronent and well-accepted classes of fngerprnt features arsng fro local rdge dscontnutes: rdge endngs and rdge bfurcatons. However, the preprocessng stage doesn t elnate all possble defects n the orgnal gray-level age and the orentaton estaton n a poor age s extreely unrelable. In order to further elnate the false nutae caused by low qualty, a nutae verfcaton echans s proposed to prove the dentfcaton per- forance. he nutae verfcaton echans uses the concept of egen-codebook to fnd the optal projecton bases for true nutae regons and false nutae regons. Experental results show that the perforance s proved effcently even less tranng data. Keywords: Fngerprnt age postprocessng, nutae-based, prncple coponent analyss, local rdge orentatons, autoatc fngerprnt dentfcaton syste Introducton A fngerprnt s the pattern of rdges and valleys on the surface of a fngertp. A total of eght dfferent types of local rdge/valley descrptons have been dentfed [], and ost of the are not used n autoatc fngerprnt dentfcaton syste (AFIS). Instead, n accordance wth the representaton of fngerprnts n the U.S. Federal Bureau of Investgaton (FBI) [], rdge endngs and bfurcatons, called nutae, are taken as the dstnctve features of the fngerprnts. he correspondng coordnate, angle, and type of the features are used to represent the fngerprnt n the atchng process. he ost pronent classes of fngerprnt atchng ethods nclude nutae-based approaches and age-based approaches [3]-[9]. Mnutae-based approaches are the ost popular ones to dentfy and verfy the fngerprnt patterns. Due to perfectons of the acqured age, the false nutae appear as the dsconnecton or sconnecton of the rdge n a fngerprnt age. ypcal approaches of nutae-prunng n the fngerprnt age postprocessng stage can be categorzed as: wthout tranng structure-based and supervsed tranng classfer-based. O Goran and Nckerson [4] calculated the statstcal characterstcs n an n-by-n atrx along the rows and coluns of the age to reove the false nutae. Xao and Raafat [3] descrbed ethods to dentfy spurous nutae and elnate the usng the structural defnton of nutae. For each nuta, statstcs of rdge length, rdge drecton and nutae drecton were used to decde the spurous nutae. Ratha et al. [6] used heurstcs based on the structural and spatal relatonshp of the nutae to valdate spurous feature ponts. Derche [8] used the nter-rdge pxel dstance as a gude to elnate these false nutae. he process of elnaton took the neatest neghborng nutae pars nto account to accurately dscard false nutae. hese researchers have proposed nutae- prunng based on wthout tranng structure or s so called rather ad-hoc technology n the postprocessng stage to elnate spurous nutae. rabhakar et al. [9] proposed a nutae verfcaton based on reexanng the gray-scale profle n a detected nutae s spatal neghborhood n the sensed age. he nutae verfcaton used the nutae classfer, whch based on supervsed tranng wth Learnng Vector Quantzaton (LVQ) [0] to learn the characterstcs of nutae n gray level age. rabhakar s algorth s easly pleented, but t has the followng drawbacks: () t needs large tranng data, () reference teplate depends on the ntalzaton and nput sequence of tranng data, (3) owng to the ult-varances of nutae, t s dffcult to fnd the representatve and

2 roceedngs of the 7th WSEAS Internatonal Conference on Multeda Systes & Sgnal rocessng, Hangzhou, Chna, Aprl 5-7, optal reference teplate for the syste. In ths paper, we use the concept of egen-codebook for the nutae verfcaton echans of fngerprnt age. We utlze the prncpal coponent analyss (CA) technology to fnd the optal projecton bases for true nutae regons and false nutae regons. hs paper s organzed as follows. In Secton II, we brefly ntroduce a sequental algorth of nutae extracton. Mnutae verfcaton echans usng CA s descrbed n Secton III. Secton IV presents experental results. Concludng rearks are gven n Secton V. Mnutae Extracton he flow chart of nutae extracton algorth s shown n Fg.. Frst, the fngerprnt age enhanceent algorth was proposed by authors [] that s based on global textural flterng and local orented averagng s used to prove the qualty of rdge-valley structure of the nput fngerprnt age. hen, a sple regonal average thresholdng (RA) schee [] s used to segent the rdge regon fro the nput age. Once the rdges are located, drectonal soothng s appled to sooth the rdges. Ratha et al. [6] use a 3 7 ask to sooth the rdge accordng to the orentaton feld of each block. After ths soothng process, soe of the lost rdges can be reconstructed and the spurous rdges can also be deleted. Before the features can be extracted, the fngerprnts have to be thnned or skeletonzed so that all the rdges are one pxel thck. In our evaluaton, the thnnng algorth proposed by Zhou et al. [3] s used because of the satsfactory perforance and the reasonable coputatonal te. Fnally, he nutae can be extracted fro the thnned age by usng the Crossng Nuber (CN) at a pont, whch s expressed as: 8 CN = 8 = () 0 = Where s the pxel value (0 or ) n a 3 3 neghborhood of, as shown n Fg.. Accordng to the characterstcs of CN, the nutae can be easly obtaned and the nutae ponts are recorded usng three-densonal feature vector, naely, the x-coordnate, the y-coordnate and the local rdge drecton θ. 3 Mnutae Verfcaton echans 3. rncpal Coponent Analyss he CA ethod uses the statstcal dstrbuton of nput saples to fnd the best projecton bases. It s wdely used n the coputer vson applcaton [4]-[6]. he advantages of CA ethod are that the prncpal egenvectors are orthogonal and represent the drectons where the sgnals have axu varaton. hs property wll speed up the convergence of odel tranng and prove the syste perforance. he CA ethod tres to fnd the projecton of the feature vector on a set of base vectors. Let X={ x t, t=, M } be a set of M n-densonal feature vectors. Hence, the covarance atrx C of X can be found as follows: M C = ( xtx)( xtx) () M t= Where M = M x t = x t (3) And M s the nuber of n-densonal feature vectors. he projecton vector, Ω = { ω, ω,..., ω }, descrbes the dstrbuton of egenvectors on -densonal lnear subspace (egenspace). hs weght s used to represent the nput nutae age and can be obtaned as follows: ωk = k ( x k x),k =, M (4) Where k s the egenvector. he transforaton atrx Φ s fored by the prncple egenvectors, ( < M) of the covarance atrx C, s the nuber of projecton bases, and the egenvectors k are organzed n Φ n such a way that all egenvectors ndces are n a descendng order correspondng to ther respectve egenvalues ( λ λ λ ). 3. Egen-Codebook Vector Quantzaton he well-known technology of VQ s called the Lnde-Buzo-Gray (LBG) algorth [7], whch s based on two necessary condtons for optalty: the centrod and the nearest neghbor condtons. In ths paper we use the odfed VQ algorth [8], whch was proposed for text-ndependent speaker dentfcaton experents. Whle saple vectors ncludng false nutae regon, rdge endngs regon and rdge bfurcatons regon belong to a code word have been processed by CA ethod, the evaluated ean vector x and egenvectors,,..., ay effectvely descrbe the dstrbuton of the saples. We propose a odfed vector quantzaton ethodology called egen-codebook vector quantzaton (ECVQ) as the dentfer for nutae

3 roceedngs of the 7th WSEAS Internatonal Conference on Multeda Systes & Sgnal rocessng, Hangzhou, Chna, Aprl 5-7, verfcaton. hs odel uses the prncpal coponent analyss (CA) ethod to evaluate the codebooks for capturng ore detals of true nutae and false nutae regon s characters. If there s a new saple xnew for classfcaton, t wll be adjusted by the ean vector and project to the orthonoral egenvectors as follows: p = ( xnew x) =,,..., (5) Where p s the projecton of x on the th new projecton base, and s the nuber of the projecton bases. Fnally, the error vector ε between xnew and the projecton of xnew on the egenspace of the code word s evaluated as follows: ε = x (6) If the nor ε s saller, the new saple ore closng to the codeword. x new p = he tranng procedure of the proposed ECVQ s descrbed as follows: Step : set the nuber of the projecton bases, nuber of code words C and axu of tranng no tes N. Step : use odfed LBG ethod to evaluate the centrods of all code words and classfes all tranng saples. Step 3: use CA ethod to evaluate the ean vector x and egenvectors,,..., for every code word. Step 4: use (6) to evaluate the error vector ε and classfy the tranng saples by ε. Step 5: If the tranng tes are less than N and saples classfcaton s not convergent, go to step 3. Step 6: store the ean vector x and egenvectors,,..., for every code word. 4 Experental Result he proposed algorth s evaluated usng NIS4 database. Saples nclude 500 fngerprnt ages. he nterested fngerprnt regon of 5 x 480 pxels wthn the orgnal age s segented and the dvded block sze n s set as 5 x 5. After capturng the requred fngerprnt regon fro the nput age, the nutae extracton algorth descrbed n sec. s appled to extract the nutae ponts. We randoly choose 0 fngerprnt age nutae for x new s tranng to label the rdge endng, rdge bfurcaton and false fngerprnt nutae by vsual nspecton and then fnd the optal projecton bases of each labelng wth CA. hrough the atrx projecton echans, the large densonal n n atrx s transfored to a sall densonal atrx. Hence, the coputaton te s largely decreased. Fg.3 shows soe exaples of the tranng saples and ts correspondng projecton bases of fngerprnt age nutae n left and rght sde, respectvely. Moreover, these fgures represent rdge endngs regon, rdge bfurcatons regon, and false nutae regon fro top to botto, respectvely. Fg.4(a) descrbes that the ntal flter echans of nutae can effectvely reove the false nutae whose are near the edge regon of fngerprnt. But t s not satsfactory result only usng the flterng echans for the nose regon of fngerprnt. After nutae verfcaton, the proposed algorth can dentfy the spurous nutae n the nose regon, but also t can antan the vald features n the clear regon, as shown n Fg.4(b). 5 Concluson In ths paper, we use the concept of egen-codebook for the nutae verfcaton echans of fngerprnt age. We utlze the prncpal coponent analyss (CA) technology to fnd the optal projecton bases for true nutae regons and false nutae regons. Experental result show that the perforance s proved effcently even less tranng data. References: [] Henry C. Lee and R. E. Gaensslen, edtors, Advances n Fngerprnt echnology, Elsever, New York, 99. [] K. Hrechak and J. A. McHugh, Autoated fngerprnt recognton usng structural atchng, attern Recognton,Vol.3, pp , 990 [3] Q. Xao and H.Raafat, Fngerprnt age processng: A cobned statstcal and structural approach, attern Recognton, Vol.4, No.0, , 99. [4] L. O Goran and J. Nckerson, An approach to fngerprnt flter desgn, attern Recognton, Vol., No., pp.9-38, 989. [5] D. C.D. Hung, Enhanceent and feature purfcaton of fngerprnt ages, attern Recognton, Vol.6, No., pp.66-67, 993. [6] N. Ratha, S. Chen, and A.K. Jan, Adaptve flow orentaton-based feature extracton n fngerprnt

4 roceedngs of the 7th WSEAS Internatonal Conference on Multeda Systes & Sgnal rocessng, Hangzhou, Chna, Aprl 5-7, ages, attern Recognton, Vol.8, No., pp , 995. [7] A.K. Jan, L. Hong, S. ankant and R. Bolle, An dentty authentcaton syste usng fngerprnts, rocessng of the IEEE, Vol.85, No. 9, pp , 997. [8] M. Derche, An algorth for reducng the effect of copresson/decopresson technque on fngerprnt nutae, Seventh Australan and Newzealand Intellgence nforaton systes Conference, 8-, pp.43-46, Nov. 00. [9] S. rabhakar, A.K. Jan, J. Wang, S. ankant, and R. Bolle, Mnuta verfcaton and classfcaton for fngerprnt atchng, IEEE. Conf. pp.5-9, 000. [0]. Kohonen, J. Kangas, J. Laaksonen, and K. orkkola, LVQ_AK: A progra package for the correct applcaton of learnng vector quantzaton algorth, n roc. Int. Jont Conf. On Neural Networks, (Baltore), pp , June 99. [] C.. Hseh, E. La, and Y.C. Wang, An effectve algorth for fngerprnt age enhanceent based on wavelet transfor, attern Recognton, Vol.36, No., 003. [] Eroglu, M. B. A khan, re-processng of Fngerprnt Iages, ECOS 97, European Conference on Securty and Detecton., 8-30, pp. 47-5, 997. [3] R. W. Zhou, C. Quek and G. S. NG, Novel sngle-pass thnnng algorth, attern Recognton. Lett., Vol.6, No., pp.67-75, 995. [4] Hurase, H., and Nayar, S., Vsual learnng and recognton of 3D objects fro appearance, Int l J. Coputer Vson, vol.4, pp. 5-4, 995. [5] Belhueur,. N., Hespanha, J.., and Kregan, D. J., Egenfaces vs. Fsherfaces: recognton usng class specfc lnear projecton, IEEE rans. attern Analyss and Machne Intellgence, vol.9, no.7, pp. 7-70, July 997. [6] Martnez, A. M., and Kak, A. C., CA versus LDA, IEEE rans. attern Analyss and Machne Intellgence, vol.3, no., pp. 8-33, Feb. 00. [7] Lnde, Y., Buzo, A., and Gray, R. M., An algorth for vector quantzer desgn, IEEE rans. Coun., vol.0, pp , 980. [8] Lee, D., Baek, S., and Sung, K.: Modfed K-eans algorth for vector quantzer desgn, IEEE. Sgnal rocessng Letters, Vol.4, No., pp. -5, 997. [9] M. urk, and A. entland, Egenfaces for recognton, Journal of Cogntve Neuroscence, vol.3, no., pp. 7-86, 99. Input fngerprnt age Enhanceent Bnarzaton Soothng hnnng Mnutae extracton Identfcaton result Fg. he flowchart of the nutae extracton algorth Fg. A 3 3 neghborhood ask of.

5 roceedngs of the 7th WSEAS Internatonal Conference on Multeda Systes & Sgnal rocessng, Hangzhou, Chna, Aprl 5-7, (a) rdge endngs regon (b) rdge bfurcatons regon (c) false nutae regon Fg.3 Soe exaples of the tranng saples and ts correspondng projecton bases of the extractng nutae. (a) (b) Fg.4 he results of (a) before and (b) after nutae verfcaton.