Multi-level dynamic error coding for face recognition with a contaminated single sample per person

Abstract

Single sample per person (SSPP) has always been a significantly challenging and practical problem for face recognition owing to limited information and facial variations. Although the existing holistic and local methods have achieved great success in face recognition with SSPP, their performances suffer a serious degradation when SSPP is accompanied by a contaminated gallery database. In this situation, the gallery set is usually disturbed by facial variations such as illumination, expression, and disguise. To solve this problem, we propose a novel method called multi-level dynamic error coding. First, a multi-level pyramid structure is constructed for holistic and local sparse representation, where gallery dictionary patches are extracted to build a local gallery dictionary and a variation dictionary is built by extracting the generic dataset patches to depict potential facial variations. Second, we further propose a scheme of dynamic error coding by constructing an error function at different levels to reduce the negative impact of variations. At last, the corrected holistic and local errors are fused to perform the classification. Experimental results on various benchmark data sets have demonstrated that our method has strong generalization ability to dictionary and is more robust against facial variations under SSPP.