Automatic Fingerprint Matching Using Extended Feature Set

Fingerprint friction ridge features are generally described in a hierarchical order at three different levels, namely, Level 1 (ridge flow), Level 2 (minutiae points) and Level 3 (pores and ridge shape, etc.). Current Automated Fingerprint Identification Systems (AFIS) generally rely only on a subset of Level 1 and Level 2 features (minutiae and core/delta) for matching. On the other hand, latent print examiners frequently take advantage of a much richer set of features naturally occurring in fingerprints. It is believed that this difference may be one of the reasons for the superior performance of fingerprint examiners over AFIS, particularly in case of difficult latent matches. Fingerprint features, other than minutiae and core/delta, are also referred to as the extended feature set (EFS). The goal of this study is to i) develop algorithms for encoding and matching extended features, ii) develop fusion algorithms to combine extended features with minutiae information to improve fingerprint matching accuracy, and iii) understand the contributions of various extended features in latent fingerprint matching. We study a number of extended features at all three levels, including ridge flow map, ridge wavelength map, ridge quality map, ridge skeleton, pores, dots, incipient ridges, and ridge edge protrusions. Feature extraction and matching algorithms are developed for each type of feature. Relative contribution of each feature towards the overall matching accuracy is evaluated by incrementally adding features to baseline features (minutiae and core/delta). The order of adding features is determined based on the amount of manual labour in feature marking and the estimated importance of features. Latent fingerprint databases, NIST SD27 and ELFT-EFS-PC, and several NIST rolled/plain fingerprint databases are used in our experiments. Based on extensive experiments, we report the following findings: i) almost all the extended features lead to some improvement in latent matching accuracy, ii) extended features at higher level are more effective in improving latent matching accuracy than those at lower level, iii) high image resolution (at least 1000 ppi) is necessary but not sufficient for reliably capturing Level 3 features. Based on our study, we would like to offer the following recommendations: i) extended features at Level 1 and Level 2 are strongly recommended to be incorporated into AFIS, ii) convenient GUI tools should be developed to help fingerprint examiners manually mark extended features (especially ridge skeleton) at Level 1 and Level 2 in latents, iii) it is crucial to improve the quality of enrolled fingerprints (so that a sufficient number of Level 3 features can be extracted) before Level 3 features can play an important role in AFIS.