Robust Label Consistent Dictionary Learning for Face Recognition

The Label Consistent K-Singular Value Decomposition (LC-KSVD) algorithm, which has been introduced recently has shown better results for learning a discriminative dictionary for recognition and classification by considering the label consistency constraint in the cost function. However this approach assumes Guassian distribution for the coding residual, which might not be true in the practical Face Recognition system due to lighting, occlusion/disguise and expression variations. In this paper, we propose a new Robust Label Consistent Dictionary Learning (RLC-DL) algorithm, which assumes that the coding residual and coefficients are identically distributed, independent and tries to find a Maximum Likelihood Estimation of the coding problem. The proposed algorithm is evaluated on the publicly available face datasets and it shows superior performance than state of the art algorithms available including LC-KSVD for face recognition.