Learning for combining fingerprint matchers: a case study FVC-OnGoing

This chapter focuses on the use of machine learning and statistical approaches to combine fingerprint matchers. The purposes of this chapter are: a) to explore the correlation among the state of the art matchers for fingerprint verification published in FVC-onGoing; b) to study different learning approaches for combining fingerprint matchers. Our aim is to investigate whether the integration of different state-of-the-art fingerprint matchers of this era can achieves performance not achievable using a single matcher, as we have already reported with the previous state-of-the-art approaches presented in FVC2004. Moreover we are interested to study the different behavior of the ensemble of matchers in the two different datasets of FVC-onGoing: a) FV-STD-1.0 contains fingerprint images acquired in operational conditions using high-quality optical scanners; b) FV-HARD-1.0 contains a relevant number of noisy images, distorted impressions, etc. Scores from the selected matchers are fused using i) statistical rules and ii) learning approaches (e.g. likelihood ratio test). The best performance, among the statistical rules, is obtained by max rule and, among the learning approaches, is obtained by a random subspace of AdaBoost of neural networks trained on features extracted from the likelihood ratio of the scores.