302IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE VO - PDF document

302IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 M anuscript received Feb. 2, 1996; revised Oct. 21, 1996. Recommended for accep-tance by B. Dom.For information on obtaining reprints of this article, please send e-mail to:transpami@computer.org, and reference IEEECS Log Number P96113.F JAIN ET AL.: ON-LINE FINGERPRINT VERIFICATION (a)(b)Fig. 1. Inkless fingerprint scanners: (a) Manufactured by Identix. (a)(b)Fig. 2. Comparison of fingerprint images captured by different meth-ods. (a) Inked impression method (from NIST database). (b) Inkless (a) (b) (c) (d) (e) (f)Fig. 3. A coarse-level fingerprint classification into six categories:properties (Fig. 3 shows a coarse-level fingerprint classifi-3 shows a coarse-level fingerprint classifi-However, no matter what type of framework is used, theclassification is based on ridge patterns, local ridge orienta-tions and minutiae. Therefore, if these properties can bedescribed quantitatively and extracted automatically from afingerprint image then fingerprint classification will be-come an easier tas

k. During the past several years, a num-ber of researchers have attempted to solve the fingerprintclassification problem [11], [3], [9], [10], [26]. Unfortunately,their efforts have not resulted in the desired accuracy. Algo-or six categories with about 90 percent classification accuracypercent classification accuracy[26]. However, to achieve a higher recognition accuracy witha large number of categories still remains a difficult problem.Fingerprint verification determines whether two finger-prints are from the same finger or not. It is widely believedthat if two fingerprints are from the same source, then theirlocal ridge structures (minutia details) match each othertopologically [11], [3]. Eighteen different types of local ridgedescriptions have been identified [11]. The two mostprominent structures are ridge endings and ridge bifurca-tions which are usually called minutiae. Fig. 4 shows ex-1) the correspondences between the template and input2) there are no deformations such as translation, rotation3) each minutia present in a fingerprint image is exactly 1) no correspondence is known beforehand,2) there are relative translation, rotation and n

onlinear3) spurious minutiae are present in both templates and4) some minutiae are missed.achieve. Fig. 5 illustrates the difficulty with an example of 304IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 Fig. 5. Two different fingerprint images from the same finger. In order5. Two different fingerprint images from the same finger. In orderIn this paper, we will introduce an on-line fingerprintverification system whose purpose is to capture finger-print images using an inkless scanner and to comparethem with those stored in the database in “real time.”block diagram of our system is shown in Fig. 6. It operates1) Off-line phase: Several impressions (depending on the2) On-line phase: The individual to be verified gives is implemented which is much faster and more reliable forminutia extraction. We propose a hierarchical approach toobtain a smooth orientation field estimate of the input fin-gerprint image, which greatly improves the performance ofminutia extraction. For minutia matching, we propose analignment-based elastic matching algorithm. This algorithmis capable of finding the correspondences betwe

en minutiaewithout resorting to an exhaustive search and has the abil-ity to adaptively compensate for the nonlinear deforma-tions and inexact pose transformations between differentfingerprints. Experimental results show that our systemachieves excellent performance in a real environment.In the following sections we will describe in detail ouron-line fingerprint verification system. Section 2 mainlytion 3 presents our minutia matching algorithm. Experi-two different inkless scanners are described in Section 4.4.the minutiae of a fingerprint is unique, and invariant withaging and impression deformations [11], [3]. This impliesthat fingerprint identification can be based on the topologi-cal structural matching of these minutiae. This reduces thecomplex fingerprint verification to minutia matching proc-ess which, in fact, is a sort of point pattern matching withthe capability of tolerating, to some restricted extent, de-formations of the input point patterns. Therefore, the firststage in an automatic fingerprint verification procedure isto extract minutiae from fingerprints. In our on-line finger-print verification system, we have implemented a minut

iaextraction algorithm which is an improved version of themethod proposed by Ratha et al. [18]. Its overall flowchartis depicted in Fig. 7. We assume that the resolution of input JAIN ET AL.: ON-LINE FINGERPRINT VERIFICATION 2.1Estimation of Orientation FieldEstimation of Orientation Fieldnew hierarchical implementation of Rao’s algorithm [17] isused. Rao’s algorithm consists of the following main steps:1) Divide the input fingerprint image into blocks of size2) Compute the gradients and at each pixel in each3) Estimate the local orientation of each block using the GijGijGijGijchchchch is the size of the block, and and are and directions, re- of the orientation field ijijijDbgbg -+360360180 represents the local neighborhood around 5); and and (tained. Fig. 8 shows the orientation field of a finger- (a) Rao’s method. (b) Hierarchical method.Fig. 8. Comparison of orientation fields by Rao’s method and the 16 and2.2Ridge Detection (on an 7 in our system), respectively. These two hxyuveuvxyuvxyv,;,cos,-=-ifctgifctg 306IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 hxyuvuvxyuvxyv,;,cos,ifctgifctg

W=-LxyLxysin,sin, the gray level values at pixel (2.3Minutia Detection denote its eight (a) Input image. (b) Orientation field. (c) Fingerprint region. (d) Ridge map. (e) Thinned ridge map. (f) extracted minutiae.Fig. 9. Results of our minutia extraction algorithm on a fingerprint im- 512) captured with an inkless scanner. (a) Input image.(b) Orientation field superimposed on the input image. (c) Fingerprint 3) orientation which is defined as the local ridge orien-4) the associated ridge.ment in the minutia matching phase. Fig. 9 shows the re-9 shows the re-matching problem is essentially intractable, features amatching problem is essentially intractable, features afidence level of each corresponding pair based on its con-sistency with other pairs until a certain criterion is satisfied.Although a number of modified versions of this algorithmhave been proposed to reduce the matching complexity[23], these algorithms are inherently slow because of theiriterative nature.The Hough transform-based approach proposed byStockman et al. [22] converts point pattern matching to aproblem of detecting the highest pe

ak in the Hough spaceof transformation parameters. It discretizes the transforma-tion parameter space and accumulates evidence in the dis-dis-chical Hough transform-based registration algorithm whichgreatly reduced the size of accumulator array by a mul-tiresolution approach. However, if the number of minutiapoint is less than 30, then it is very difficult to accumulateenough evidence in the Hough transform space for a reli-able match.Another approach to point matching is based on energyminimization. This approach defines a cost function basedon an initial set of possible correspondences and uses anappropriate optimization algorithm such as genetic algo-rithm [1] and simulated annealing [21] to find a possiblesuboptimal match. These methods tend to be very slow andare unsuitable for an on-line fingerprint verification system.In our system, an alignment-based matching algorithm isimplemented. Recognition by alignment has received agreat deal of attention during the past few years [12], be-cause it is simple in theory, efficient in discrimination, andfast in speed. Our alignment-based matching algorithmdecomposes the minutia matching into two stages

:1)Alignment stage, where transformations such astranslation, rotation and scaling between an input anda template in the database are estimated and the inputminutiae are aligned with the template minutiae ac-3.1Alignment of Point Patterns ridges (see Fig. 10). During the minutiae detection stage, and denote the sets of ridges asso- 308IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 19971) For each ridge is the minimal length of the two ridges and represent the distances from point on the and to the x-axis, respectively. The sam- (0 1) is, then go to step 2,ridges (Fig. 10). Generally, a least-square method can be- and ( are the and coordi- and between the two ridges is the minimal length of the two ridges and and are radial angles of the and 3) Denote the minutia ( inputcossinsincos001 represents the corre-3.2Aligned Point Pattern MatchingAligned Point Pattern Matchingfactory performance in practice, because local deformationsmay be small while the accumulated global deformationscan be quite large. We have implemented an adaptive elas-tic matching algorithm with the ability to compensate th

eminutia localization errors and nonlinear deformations.LetPxyxyPPP111,,,...,,,ejejQxyxyQQQ111,,,...,,,ejej minutiae in the input image which is1) Convert each minutia point to the polar coordinate xxyy-+-ejejiii*** are the coordinates of a minutia, are the coordinates of the reference minu- is the representation of the minutia represents the radial represents the radial angle, and repre-2) Represent the template and the input minutiae in thePrerePPP111,,,...,,,ejejQrereQQQ111,,,...,,,ejejPPP*** and QQQ*** represent the cor-3) Match the resulting strings and with a dynamic- with a dynamic-betweenPp and Qp which is described below.4) Use the edit distance between and to establish and max,,porating an elastic criteria into a string matching algorithm.Generally, string matching can be thought of as the maxi-mization/minimization of a certain cost function such asthe edit distance. Intuitively, including an elastic term inthe cost function of a string matching algorithm can achievea certain amount of error tolerance. Given two strings PpandQp of lengths and CmnCmnCmnCmnwmnafaf--+000 wmnrre-++abgqaaee=-+360360180=-+360360180 are the weights asso

ciated with each specify the bounding is a pre-specified penalty for a mismatch. Suchby adjusting the bounding box (Fig. 11) when an inexact 310IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 -++wmnrreabgqddhmnmnr++=+afafddhmnmnr++=+afafmnmne++=+afafmnmne++=+afaf and = 30; = 1.0; = 2.0; = 0.1; = 200( + + = 0.5. The values of images. Fig. 12 shows the results of applying the matching Fig. 11. Bounding box and its adjustment. (a) (b) (c) (d)Fig. 12. Results of applying the matching algorithm to an input minutia 380. Set 2 con- 480. When these fingerprint images were cap- ages in our database. Approximately 90 percent of the fin-10 percent of the fingerprint images in our database are notof good quality (Fig. 15), which are mainly due to large15), which are mainly due to large1) we concentrate on live-scan verification, and2) NIST-9 fingerprint database is a very difficult finger- Fig. 13. Fingerprint images captured with a scanner manufactured by 380; all the three images are Fig. 14. Fingerprint images captured with a scanner manufactured by 480; all the three 4.1Matching 1

79) matchings have been performed on test 609) matchings on test Set 2. Theshown in Fig. 16. It can be seen from this figure that therethe threshold value. As we have observed, both the incor-images with poor quality such as those shown in Fig. 15. We (a) Identix Fig. 16. Distributions of correct and incorrect matching scores; vertical 312IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 19974.2Verificationtion is established. With this scheme, a 100 percent verifica-16 percent on both test sets. Again, this reject rate can bepractice, using a k-nearest neighbor type of matching is ade-few seconds. Table 3 shows the CPU requirements of ouron a SPARC 20 workstation. It indicates that our on-line 2099.839 %11.23 %2099.426 %11.23 % 2299.947 %13.33 %2299.863 %14.55 % 2499.984 %16.48 %2499.899 %16.78 % 2699.994 %20.49 %2699.969 %20.20 % 2899.996 %25.19 %2899.989 %23.15 % 30100 %27.72 %3099.999 %27.45 % (a) Using Identix system (180 images). (b) Using Digital Biometrics system (610 images).Matching RateNumber of 191.17 %192.13 % 294.72 %294.40 % 396.89 %397.06 % 498.17 %497.67 % 598.89 %598.44 % 699.39 %6

99.11 % 799.72 %799.70 % 899.83 %899.79 % 999.94 %999.91% (a) Using Identix system (180 images). (b) Using Digital Biometrics System (610 images).TABLE 3AVERAGECPU TIME FOR MINUTIAEXTRACTIONATCHING ON A ORKSTATIONMinutia Extraction(seconds)Minutia Matching(seconds) 5.352.557.90 [1] N. Ansari, M.H. Chen, and E.S.H. Hou, “A Genetic Algorithm forPoint Pattern Matching,” Chapt. 13, B. Soucek and the IRIS[2] P.E. Danielsson and Q.Z. Ye, “Rotation-Invariant Operators Ap-[3] Federal Bureau of Investigation, [4] T.H. Cormen, C.E. Leiserson, and R.L. Rivest, [5] D.C.D. Hung, “Enhancement and Feature Purification of Finger-[6] S. Gold and A. Rangarajan, “A Graduated Assignment Algorithm[7] L. O’Gorman and J.V. Nickerson, “An Approach to Fingerprint[8] K. Karu and A.K. Jain, “Fingerprint Registration,” Research Re-[9] K. Karu and A.K. Jain, “Fingerprint Classification,” [10] M. Kawagoe and A. Tojo, “Fingerprint Pattern Classification,”[11] H.C. Lee and R.E. Gaensslen, eds., [12] D.P. Huttenlocher and S. Ullman, “Object Recognition Using[13] Z.R. Li and D.P. Zhang, “A Fingerprint Recognition System With[14] L. Coetzee and E.C. Botha, “Finger

print Recognition in Low[15] B. Miller, “Vital Signs of Identity,” vol. 31, no. 2,[16] A. Ranade and A Rosenfeld, “Point Pattern Matching by Relaxa-[17] A.R. Rao, [18] N. Ratha, S. Chen, and A.K. Jain, “Adaptive Flow Orientation[19] A. Sherstinsky and R.W. Picard, “Restoration and Enhancement[20] D.B.G. Sherlock, D.M. Monro, and K. Millard, “Fingerprint En-[21] J.P.P. Starink and E. Backer, “Finding Point Correspondence Us- vol. 28, no. 2,[22] G. Stockman, S. Kopstein, and S. Benett, “Matching Images to[23] J. Ton and A.K. Jain, “Registering Landsat Images by Point vol. 27,[24] V.V. Vinod and S. Ghose, “Point Matching Using Asymmetric vol. 26, no. 8, pp. 1,207-[25] C.I. Watson and C.L. Wilson, [26] C.L. Wilson, G.T. Gandela, and C.I. Watson, “Neural-Network[27] Q. Xiao and Z. Bian, “An Approach to Fingerprint Identification 314IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 19, NO. 4, APRIL 1997 received a BTech degree in 1969 (1990–1994), and he also and the (Prentice Hall, 1988), has (Springer- (Springer-Verlag, 1989), (Academic (Elsevier, 1993). He is a received the BS and MS degrees in (S‘82-M’84-F’96) received th