Robust low-rank image representations by deep matrix decompositions

A novel approach based on low-rank representations (LRRs) for image representations is proposed. LRR seeks the lowest-rank representations among all the coefficient matrices that represent the images as linear combinations of the basis in the given dictionary. Unlike LRR methods of enforcing additional constraints on the representation and dictionary, an iterative process in which the low-rank decomposition is performed on the coefficient matrices has been developed. The rank of the representation matrices will be lower and lower with the iterations, termed as the deep low-rank (DLR) method. Extensive experiments were conducted to verify the state-of-the-art performance for classification tasks of the DRL method.