Face recognition using kernel collaborative representation and multiscale local binary patterns

Abstract

The role of collaboration between classes is a key to capture discriminative information among the different classes of image samples and can lead to very good and robust recognition rates. One of the modern approaches that make full use of the collaboration among classes in defining the query sample is Collaborative Representation with regularized least square (CRC-RLS). But, it uses the image intensity features to represent the template and query images and is susceptible to changes in illumination and alignment of cropped faces. Local binary patterns (LBP) have emerged as a very powerful discriminative texture descriptor in representing images and are widely used in state-of-the-art algorithms in terms of acquiring higher accuracies as compared to other descriptors. Multi-scale LBP extends the local binary pattern to multi-scale representation. In this chapter, multi-resolution LBP is employed with CRC-RLS to handle problems associated with face recognition such as illumination, occlusion, disguise, etc. In addition, a kernel version of CRC-RLS is also proposed. The efficacy of our proposed KCRC-RLS is evaluated on four challenging image databases, i.e., AR, Extended Yale B, FERET and ORL capturing a different set of problems including illumination, gesture, occlusion, small pose, etc. The results on all databases indicate a significant improvement in comparison to leading approaches like linear regression (LRC), sparse representation (SRC).