Face recognition technology robust to pose variation and partial occlusion

Abstract

In recent years, numerous studies have attempted to develop algorithms for unconstrained face recognition. However, the problem of addressing a partial face, which does not contain whole facial information, has not yet been extensively explored. Because a holistic face includes information of semantic correspondences among human faces, it is frequently used to align a face with an arbitrary pose. In this study, we propose a partial face recognition method that does not require face alignment and ensures robustness to variations in facial poses. The proposal is based on the idea that a partial face also contains a sufficient amount of distinctive features even though pose changes occur. Accordingly, we present an affine-simulation-based patch representation method, which covers pose variations, as well as a classification scheme to integrate local identity information from the partial face area. We verified the efficacy of the proposed algorithm in experiments conducted on the AR, Extended Yale B, and Multi-PIE databases. The results show the effectiveness of the proposed algorithm and its robustness against partial occlusion and pose variations. In addition, we develop a coarse head pose estimator as a preprocess step of face recognition technologies. Head pose estimation continues to be a challenge for computer vision systems because extraneous characteristics and factors that lack pose information can change the pixel values in facial images. Thus, to ensure robustness against variations in identity, illumination conditions, and facial expressions, we propose an image abstraction method and a new representation method (local directional quaternary patterns, LDQP), which can remove unnecessary information and highlight important information during facial pose classification. We verified the efficacy of the proposed methods in experiments, which demonstrated its effectiveness and robustness against different types of variation in the input images. In this dissertation, we address the two proposed approaches respectively, where they can be utilized as a complementary part to each other.