Locality Features Encoding in Regularized Linear Representation Learning for Face Recognition

Abstract

Regularized linear regression based representation techniques for face recognition (FR) have attracted a lot of attention in past years. The l1-regularized sparse representation based classification (SRC) method achieves state-of-the-art results in FR. However, recently several studies have shown the role of collaborative representation (CR) that plays a crucial role for the success of SRC in robust classification and not the l1-regularization constraints on representation. In this paper, we propose a novel Robust Locality based Collaborative Representation (RLCR) method using weighted regularized least square regression approach that incorporates the locality structure and feature variance among data elements into linear representation. RLCR is an extension of collaborative representation based classification (CRC) approach, a recently proposed fast alternative to SRC. The performance of CRC method dramatically decreases when the feature dimension is low or the number of training samples per subject is limited. RLCR improves classification performance over that of original CRC formulation. Experimental results on real world face datasets using low dimensional as well as high dimensional linear feature space have demonstrated the effectiveness of the proposed method and is found to be very competitive with the state-of-the-art image classification methods.