In the field of face recognition, Sparse Representation (SR) has receivedconsiderable attention during the past few years. Most of the relevantliterature focuses on holistic descriptors in closed-set identificationapplications. The underlying assumption in SR-based methods is that each classin the gallery has sufficient samples and the query lies on the subspacespanned by the gallery of the same class. Unfortunately, such assumption iseasily violated in the more challenging face verification scenario, where analgorithm is required to determine if two faces (where one or both have notbeen seen before) belong to the same person. In this paper, we first discusswhy previous attempts with SR might not be applicable to verification problems.We then propose an alternative approach to face verification via SR.Specifically, we propose to use explicit SR encoding on local image patchesrather than the entire face. The obtained sparse signals are pooled viaaveraging to form multiple region descriptors, which are then concatenated toform an overall face descriptor. Due to the deliberate loss spatial relationswithin each region (caused by averaging), the resulting descriptor is robust tomisalignment & various image deformations. Within the proposed framework, weevaluate several SR encoding techniques: l1-minimisation, Sparse AutoencoderNeural Network (SANN), and an implicit probabilistic technique based onGaussian Mixture Models. Thorough experiments on AR, FERET, exYaleB, BANCA andChokePoint datasets show that the proposed local SR approach obtainsconsiderably better and more robust performance than several previousstate-of-the-art holistic SR methods, in both verification and closed-setidentification problems. The experiments also show that l1-minimisation basedencoding has a considerably higher computational than the other techniques, butleads to higher recognition rates.
展开▼
