Making better biometric decisions with quality and cohort information: A case study in fingerprint verification

Automatically recognizing humans using their biometric traits such as face and fingerprint will have very important implications in our daily lives. This problem is challenging because biometric traits can be affected by the acquisition process which is sensitive to the environmental conditions (e.g., lighting) and the user interaction. It has been shown that post-processing the classifier output, so called score normalization, is an important mechanism to counteract the above problem. In the literature, two dominant research directions have been explored: cohort normalization and quality-based normalization. The first approach relies on a set of competing cohort models, essentially making use of the resultant cohort scores. A well-established example is the T-norm. In the second approach, the normalization is based on deriving the quality information from the raw biometric signal. We propose to combine both the cohort score- and signal-derived information via logistic regression. Based on 12 independent fingerprint experiments, our proposal is found to be significantly better than the T-norm and two recently proposed cohort-based normalization methods.