Sparse Representation for Face Recognition

Robust face recognition via sparse representation is a technique developed for a face recognition system, where a rich set of carefully controlled training face images are provided. Under the assumption that all the training samples from a single class lie in a low-dimensional subspace of a high-dimensional space, this technique tries to code the given test face image as a sparse linear combination of all the training images themselves, i.e., use the fewest possible training samples to interpret the test sample. Sparse representation-based classification (SRC) measures the sparsity of the coding vector by l0-norm, which counts the number of nonzero entries. Since the l0-minimization problem is NP-hard, the l1-minimization, as the closest convex function to l0minimization, is employed to find the sparsest coding vector. By optimizing the sparsity of such an over-complete linear representation, the dominant nonzero entries in the coding vector can reliably indicate the identity of the test sample. Finally, SRC performs the classification by checking which class yields the minimum representation error. This technique can effectively handle errors due to occlusion and corruption uniformly by exploiting the sparsity on the location of the distorted pixels in the face image.