PCA filter based covariance descriptor for 2.5D face recognition

Region covariance matrix (RCM) as a feature descriptor is shown promising in various object detection and recognition tasks. However, vanilla RCM breaks down in face recognition due to its inadequacy in extracting discriminative features from facial image. In this paper, cascaded Principle Component Analysis (PCA) filter responses that derived from the multi-layer PCA network are leveraged to extract the sufficient discriminative facial feature for RCM construction. The factors that affect the performance of cascaded PCA filter responses in forming RCM for 2.5D face recognition is investigated. To be specific, the influence of patch size and filter numbers of cascaded PCA filter responses to RCM is probed. Besides that, block division is proposed for RCM to further enhance the accuracy performance. Experimental results have demonstrated the efficacy of the proposed approach.