Analysis of unsupervised template update in biometric recognition systems

Abstract

Performance of mono- and multi-modal biometric systems depends on the representativeness of enrolled templates. Unfortunately, error rate values estimated during the system design are subject to variations due to several aspects: intra-class variations arising on small-medium time-window, and ageing, which is the natural process involving any biometrics. This causes the increase of the False Rejection Rate (genuine users are no more recognized) or the False Acceptance Rate (impostors are misclassified as genuine users), or both. In fact, several vendors strongly suggest to repeat enrolment sessions in order to collect, over time, a set of templates representative enough. As alternative, automatic template update algorithms, which exploit the own-knowledge of the mono- or multi-modal biometric system, on a batch of samples collected during system operations without the human supervision, have been proposed.Preliminary experimental results have shown that these algorithms are promising, but the motivation of their behaviour has not yet been explained. This paper is aimed to fill such gap, by showing that behaviour of self- and co-update may be explained by exploiting the concept of path-based clustering. Therefore, problems as ‘intra-class’ variations and ageing are dependent on the path-based cluster followed by each algorithm. Moreover, we show that the performance of co-update is superior than that of self-update, by a simulative model. The path-based clustering theory applied to self- and co-update algorithms, as well as the proposed model, are experimentally validated on the large DIEE Multimodal data set, the only one publicly available and explicitly conceived for comparing template update algorithms.