Discriminative patch-based sparse representation for face recognition
Wenming Yang, Riqiang Gao, Ying Xu, Xiang Sun, Qingmin Liao · 2016
This paper illustrates that elaborately selecting some discriminative patches instead of using all patches is helpful for recognition tasks. We propose an effective method called bagging greedy algorithm (BGA) for choosing a collection of patches that is discriminative in terms of classification. Compared with greedy algorithm, BGA is less likely to get into local optimum. Based on the subset of patches, discriminative patch based sparse representation classification (DPSRC) is presented, in which the corresponding patches are concatenated and the reconstruction error are normalized by l1norm. Compared with state-of-arts, our method shows advantages especially for non-occlusion cases.