Fingerprint Classification and Indexing

5.5 Summary

Fingerprint classification has been the subject of several pattern recognition studies over the last three decades. Different solutions have been proposed and it is now possible to design classification systems that are able to meet the FBI requirement of 99% accuracy with a maximum rejection of 20%. However, it is unlikely that exclusive classification would make it possible to significantly reduce the effort of searching for a single fingerprint in the absence of other information (e.g., sex, age, race, etc.). Continuous classification and related indexing/retrieval strategies based on features extracted from fingerprints seem to be more promising alternatives for efficient implementations of the identification task in a variety of applications.

In AFIS and other semi-automatic civilian applications, the enrollment is supervised, the quality of the input fingerprint can be checked, and manual intervention is possible to correct feature extraction errors: this allows us to design indexing/retrieval mechanisms that achieve a relatively small retrieval error and a good penetration rate. On the other hand, the identification task in a completely automatic system, working with live-scan dab images, has more severe constraints: database templates and input fingerprint images are often low quality and provide only a partial impression of the finger, and the system response is usually expected within a few seconds. The development of such an automatic fingerprint-based identification system for large databases is a challenging task, due to both accuracy (see subsection “Identification System Errors” in Section 1.4) and speed issues. Multimodal systems (see Chapter 7), seem to be the most promising way to improve accuracy (De Boer, Bazen, and Gerez, 2001), and to derive sequential approaches that progressively refine the search when a large database has to be explored.