Context-Sensitive Self-Updating for Adaptive Face Recognition

Abstract

Performance of state-of-the-art face recognition (FR) systems is known to be significantly affected by variations in facial appearance, caused mainly by changes in capture conditions and physiology. While individuals are often enrolled to a FR system using a limited number of reference face captures, adapting facial models through re-enrollment or through self-updating with highly confident operational captures has been shown to maintain or improve performance. However, frequent re-enrollment and updating can become very costly, and facial models may be corrupted if misclassified face captures are used for self-updating. This chapter presents an overview of adaptive FR systems that perform self-updating of facial models using operational (unlabelled) data. Adaptive template matching systems are first revised, with a particular focus on system complexity control using template management techniques. A new context-sensitive self-updating approach is proposed to self-update only when highly confident operational data depict new capture conditions. This allows to enhance the modelling of intra-class variations, while mitigating the growth of the system by filtering out redundant information, thus reducing the need to use costly template management techniques during operations. A particular implementation is proposed, where highly confident templates are added according to variations in illumination conditions detected using a global luminance distortion measures. Experimental results using three publicly available FR databases indicate that this approach enables to maintain a level of classification performance comparable to standard self-updating template matching systems, while significantly reducing the memory and computational complexity over time.