Robust Face Recognition Via Dual Nuclear Norm Low-Rank Representation And Self-Representation Induced Classifier

For robust face recognition, we particularly focus on the ubiquitous scenarios where both training and testing images are corrupted due to occlusions. In the previous low-rank based methods, each error image is stacked into a vector and formed an error matrix together, and then L1-norm or L2-norm is utilized to characterize the matrix. However, the structure information of error images will lose in the step of stacking. In this paper, we propose a novel method by utilizing a low rank hypothesis on the representation term and the error term simultaneously. For classification, we also adopt the discriminative self-representation induced classifier with more effectiveness and efficiency. Experimental results on different face recognition tasks show that our proposed method achieves comparable or superior performance to sonic state-of-the-art methods.