Learning robust and discriminative low-rank representations for face recognition with occlusion

For robust face recognition tasks, we particularly focus on the ubiquitous scenarios where both training and testing images are corrupted due to occlusions. Previous low-rank based methods stacked each error image into a vector and then used L1 or L2 norm to measure the error matrix. However, in the stacking step, the structure information of the error image can be lost. Depart from the previous methods, in this paper, we propose a novel method by exploiting the low-rankness of both the data representation and each occlusion-induced error image simultaneously, by which the global structure of data together with the error images can be well captured. In order to learn more discriminative low-rank representations, we formulate our objective such that the learned representations are optimal for classification with the available supervised information and close to an ideal-code regularization term. With strong structure information preserving and discrimination capabilities, the learned robust and discriminative low-rank representation (RDLRR) works very well on face recognition problems, especially with face images corrupted by continuous occlusions. Together with a simple linear classifier, the proposed approach is shown to outperform several other state-of-the-art face recognition methods on databases with a variety of face variations.