15

80 70

GAR (%)

40 30 20 10 0 0 0.01 0.02 0.03 0.04 0.05

FAR (%)

2 < 100; threshold to nd the corresponding points, Td = 12; threshold to nd the t corresponding triangles, Tn = 8. Figures 15.12 and 15.13 show identi cation results based on classi cation and indexing, respectively. Note that GAR cannot reach 100.0%. One important reason is that bad-quality images do not provide enough similarity information to be used in veri cation, and the NIST-4 database is a very dif cult database. Using the

80 70 60 50

GAR (%)

p = 5%

40 30 20 10 0 0 0.02 0.04

FAR (%)

classi cation-based approach, GAR is 77.2% when FAR is 4.1 10 2 %, while using the indexing-based approach with p = 5%, GAR is 77.2% and FAR is 8.0 10 3 %. It shows that in order to achieve similar GAR in identi cation, we only need to search 5% of the database by indexing-based approach for identi cation, while classi cationbased approach for identi cation may need to search 20% of the entire search space. FAR for indexing-based approach is much less than that for the classi cation-based approach. The classes R, L, W, A, and T are uniformly distributed in NIST-4. However, in nature, the frequencies of their occurrence are 31.7%, 33.8%, 27.9%, 3.7%, and 2.9%, respectively. So, using the classi cation-based approach the search space that needs to be searched will be more than 30.0%, since there are fewer ngerprints that belong to A and T classes in nature than to other classes.

In this chapter, we compared the performance of two approaches for identi cation. One is the traditional approach that rst classi es a ngerprint into one of the ve classes (R, L, W, A, T) and then performs veri cation. The alternative approach is based on indexing followed by veri cation. Using state of the art highly competitive approaches for classi cation, indexing, and veri cation, we compared the performance of the two approaches for identi cation using the NIST-4 ngerprint database. We found that the indexing technique performs better considering the size of search space (5% versus 20%) that needs to be examined. Also, for the same GAR (77.2%) the FAR performance (8.0 10 3 % versus 4.1 10 2 %) of indexing-based approach is lower. Thus, the indexing based approach provides a potential alternative to the traditional classi cation-based approach commonly used for ngerprint identi cation. Also it is possible to use the indexing approach within each of the classes after the classi cation has been done. This will expedite the identi cation performance of a classi cation-based approach.

1. X. Tan, B. Bhanu, and Y. Lin, Fingerprint identi cation: classi cation vs. Indexing, in IEEE International Conference on Advanced Video and Signal-Based Surveillance, Miami, FL, July 21 22, 2003, pp. 151 156. 2. K. Karu and A. K. Jain, Fingerprint classi cation, Pattern Recognit. 29(3):389 404, 1996. 3. R. Cappelli, A. Lumini, D. Maio, and D. Maltoni, Fingerprint classi cation by directional image partitioning, IEEE Trans. Pattern Anal. Mach. Intell. 21(5):402 421, 1999. 4. R. Cappelli, D. Maio, and D. Maltoni, Fingerprint classi cation based on multi-space KL, in Proceedings of the Workshop on Automatic identi cation Advances Technologies, 1999, pp. 117 120. 5. Y. Yao, G. L. Marcialis, M. Pontil, P. Frasconi, and F. Roli, Combining at and structured representations for ngerprint classi cation with recursive neural networks and support vector machines, Pattern Recognit. 36(2):397 406, 2003. 6. C. I. Watson and C. L. Wilson, NIST special database 4, ngerprint database, U.S. National Institute of Standards and Technology, 1992. 7. R. S. Germain, A. Califano, and S. Colville, Fingerprint matching using transformation parameter clustering, IEEE Comput. Sci. and Eng. 4(4):42 49, 1997.