Can the Surveillance System Run Pose Variant Face Recognition in Real Time?

Abstract

This paper presents an approach for face recognition across pose variations when only one sample image per person is available. From a near frontal face sample image, virtual views at different off-frontal angles were generated and used for the system training task. The manual work and computation burden, thus, are put on the offline training process, that makes it possible to build a real-time face recognition surveillance system. Our work exploited the inherent advantages of "single" HMM scheme, which is based on an ID discrete hidden Markov model (ID-DHMM) and is designed to avoid the need of retraining the system whenever it is provided new image(s). Experiment results on the CMU PIE face database demonstrate that the proposed scheme improves significantly the recognition performance