Facial Biometric Template Post-processing by Factorization

Facial recognition systems have reached a relatively mature stage nowadays. In a typical facial recognition system, an input facial image is converted into a biometric template that is matched against stored templates in the database to have a final decision. This paper addresses a post-processing step applied to the biometric template in order to improve the overall system's recognition accuracy. The biometric template is built upon the well known Local Binary Pattern (LBP) method that is considered one of the state-of-the-art approaches for feature extraction in the field. As novel step, we have included a postprocessing module the template that decomposes the template into factors via Non-negative Matrix Factorization (NMF). By so doing we keep only the discriminative information while disregarding irrelevant information. By projecting the initial template into the NMF factors, we showed that the accuracy is increased compared to the case where this step is skipped. The method is applied to a challenging facial database, namely, the extended Yale Face Database B, faces acquired under difficult illumination conditions.