High-dimensional face data separation for recognition via low-rank constraints

Sparse and low-rank modeling have been proved to be promising techniques for visual understanding. Based on the methodology, this paper proposes a novel method for robust face recognition, where both the training and test samples might contain corruption or occlusion. In the method, illumination model and low rank matrix recovery with structural incoherent between different training classes are united for separating discriminant low-rank identification information matrix and error matrix, based on which a sparse and dense combined representation of corrupted test sample is calculated. The representation together with the two parts of information dictionaries are utilized for the final identification of test sample. The experimental results show the effectiveness of the method.