Robust face signature and adaptive matching of face sequences for video face recognition

Abstract

Video face recognition (FR) is of considerable interest to applications such as video surveillance and video search. Particularly, FR for the purpose of tagging online video content is attracting significant attention, given the vast amount of video content created and consumed by end-user. In this paper, we present a new video FR method based on face sequence matching. Specifically, we designed a face signature robust to variations in illumination and facial pose, by making use of face pose clustering and weighted feature fusion. The use of face signature and adaptive distance metric allows for effective face sequence matching. We evaluated the proposed FR method using the VidTIMIT video database and a database consisting of YouTube video clips, adopting experimental settings encountered by real-world video FR applications. Our experimental results show that the proposed FR method is able to achieve high recognition accuracy, meeting the requirements of practical applications.