Face Recognition Via Patch-based Sparse Representation

The paper sloves the problem of face recognition with single training sample per person?SSPP?.It is diffcult to predict the varations of query samples by limited training sample.Especially,many traditional face recognition method will deteriorate significantly or fail to work when met with the problem of SSPP.This paper will designed robust classification model to avoid the effect of facial varations,and decide how to incorporate global facial information into local face model.The theroy of sparse representation is applied into local methods,as follows:1?Each image is represented by a collection of face patches from intra-class variance dictionaries and gallery images,and a group sparsity constraint will be imposed on reconstruction coefficients byl_2,1 norm constraint.The coefficients solution will harvest special structured information and be more discriminative.Considering that all patches from same people have similar sparse coding,non-informative facial region will be reconstructed by the right people.The method overconmes the advantage of global image representation which avoid the effect of less discriminative patches,and harvest the advantage of local matching method to avoid the side effect of those patches which are severely corrupted by variance.In order to decrease difference between query face and training sample,dictionary learning method which sparse representation coefficients have a specific structure is proposed to decrease the impact of intra-class variance.2?Instead of holistic method,local methods can make full use of dicriminative face regions.The paper present a measure for the importance of local facial regions,and the importance of different regions can be decided by the weight.The 8 closet neighboring patches are extracted to address the misalignment of a patch.The query image can be well represented by intra-class variance dictionary and gallery sample.The weights will be calculated according to reconstruction residual and the significance of different regions.In order to acquire more variance information,the discrimative dictionary learning method which dictionary atoms are inconsistent is proposed to represent variance information as much as possible.