On the Stability of Ranks to Low Image Quality in Biometric Identification Systems

摘要

The goal of a biometric identification system is to determine the identity of the input biometric probe. This is accomplished using a matcher which compares the input probe data against each labeled biometric data present in the gallery database. The output is a set of similarity scores that are ranked in decreasing order. The identity of the gallery entry corresponding to the highest similarity score (i.e., rank 1) is associated with that of the probe. In multibiometric systems, the outputs of multiple biometric matchers are combined. Such a combination, or fusion, can be accomplished at the score level or rank level (apart from other levels of fusion). In the literature, rank is believed to be a stable statistic. However, this belief has not been experimentally demonstrated. The contribution of this paper is to investigate the stability of ranks to the image quality degradation in both unimodal and mul-timodal scenarios. Experiments were carried out using two databases: 1) West Virginia University (WVU) dataset, composed of four fingerprints per subject for 240 subjects, 2) Face and Ocular Challenge Series (FOCS) collection, composed of three frontal faces per subject for 407 subjects. Experimental results show that, in a unimodal scenario when dealing with low quality data, ranks are more stable than scores. However, such a rank stability is not verified when fusing multiple matchers. Experiments demonstrate that, in the presence of low quality data, performance achieved by score-level fusion is better than that one achieved by rank-level fusion.