Cohort based Score Normalization

Automatic biometric authentication has long been an active research filed driven by its wide range of pratical applications. An authentication framework typically involves two stages: the enrollment stage (bulding a template model for each user) and the test stage (validating the authenticity of a query sample to its claimed identity). There are many degrading factors such as noise of biometric signals which can affect the matching procedure between a query sample and the claimed model, thereby making it unreliable to directly use the original matching score for the final decision making. Since making a classifier be capable of removing all the degrading effects is difficult, post-processing the original matching score turns out to be an important issue in biometric authentication, i.e. score normalization. In the task of normalizing against score distribution variations between the enrollment and query sessions, measuring the degradation effect in relation to a pool of impostors of the claimed identity is sensible and useful because both the claimed model and its impostor are subject to the same degradation. These impostor of the claimed template are known as cohort models. By either estimating the score distribution parameters or exploiting discriminative patterns from the sorted cohort scores, score distribution variations presented in different classes can be well measured, therefore making it more credible the final decision making.