Face Images Recognition Based On Discrimination Dictionary Learning

Face recognition has been one of the popular research hotspot, and there are numerous and widespread use in human life and business development base on face recognition. Recently, sparse representation based classification has been successfully used in face recognition, and got wide attention. How to learn a robust discriminative dictionary to classify the samples quickly has been a burning question. Based on analysis and summary the study work at home and abroad, this paper discuss the improved method of sparse representation based classification for face recognition in the following three aspects.Firstly, considering the key features of facial images and structure information of each subjects effect on the recognition. This paper proposed a dimensionality reduction based on structured sparse representation method for face recognition. To begin with, using training samples in each subject to learn the dimensionality reduction mapping matrix by unsupervised learning model, and then mapping all training samples on the matrix to get the dictionary. At last, classify the testing samples by structured sparse representation.Secondly, this paper proposed support vector guided dictionary learning based on Gabor feature. Firstly, considering the Gabor feature is robust to the change of illumination, poses and expressions, the paper extract the Gabor feature of face images with multi-scales and multi-orientation; and then deriving the formula of discriminative term in the discriminative dictionary, once again, design the adaptive weight learning modeling, this paper learning a new discriminative dictionary learning modeling and the same time learning the parameter of classifier, at last, the testing samples will be classified by the new classifier.Thirdly, considering problem of the small sample per person in the training samples, this paper proposed multi scales patch based collaborative representation based on statistical local feature. First of all, partitioning the images into multi-scale patch, and using the max pooling method to extract statistical local feature on each scale; and then classify each patch in the patch based collaborative representation modeling, there will be different recognition result on each scale; this paper learns weight values of each scale by margin distribution optimization combine with l1-regularized;at last, summing all the multi-scale result to get the final result.