Score-based fusion using quality measures in a semi-supervised identity verification system

The performance of a biometric verification system is affected by how good each user is represented in the user gallery. Due to the infinite number of pose variations, illumination changes and other intra-class variations in the biometric samples, it is impossible to collect all variations in a totally supervised manner. Adaptive biometric systems that use semi-supervised learning techniques are suggested recently in the literature to continuously update user galleries during the system operation. In this work we propose a method for self-training in a bi-modal biometric verification system by making use of the unlabeled data collected during system operation. The novelty of the proposed approach is the use of quality measures to assign weights to individual matchers in a dynamic way and to use the quality information for updating the user gallery. Preliminary results show that using quality measures in the fusion process can increase the accuracy of verification over time, particularly when the percentage of degraded input patterns is substantial.