Face Recognition Based On Sparse Representation

Face recognition technology is a hot subject in the field of image processing, with the improvement of compressed sensing theory, the sparse representation is successfully used for face recognition. Compared with the traditional method of face recognition, the classification method based on Sparse Representation (SRC) saves the steps of feature extraction, and has good robustness to noise, light and shade. Currently, research based on SRC is very hot. In this paper, related research based on SRC also has been carry on. The main work is as follows:(1) Studied in feature extraction, although classification based on sparse representation is don’t need to extract features, but the typical characteristics are sure to better characterization of samples, so this paper will combine Gabor transform and another mature feature extraction method PCA, Gabor can extract the texture features of different scales and directions, PCA can eliminate redundant information from these texture feature, and obtain more refined features.(2) The theory of compressed sensing and sparse representation is studied in this paper, in view of the problems such as negative coefficient and coefficients’sparse is not good enough, which appeared in solve the sparse coefficient of traditional orthogonal matching pursuit algorithm and database function. Then this paper proposed an orthogonal matching pursuit algorithm based on immunology which can solve these problems. Lots of experiments proved that the proposed method id very good to solve the problems that exist in traditional method.(3) Dictionary learning framework is introduced in this paper, and studied the Metaface, K-SVD and the expansion of K-SVD, the introduction of dictionary learning framework subverted the traditional conception of dictionary. Compared with the traditional dictionary, the number of atoms of the learned dictionary is greatly reduced, and the projection of the samples in the new dictionary is more sparse, and improve the recognition performance.(4) Further studied the kernel sparse representation classification algorithm(KSRC), this algorithm through kernel technology to map samples into high dimension space, use sparse representation in the new space can change the distribution of samples, and test samples get better sparse reconstruction by a linear combination of the training samples from the same class. The non-zero elements of sparse coefficient will closer together with training samples which from the same class with the test sample, so as to improve the recognition performance.Feature extraction, dictionary learning and classification recognition was studied in this paper, the combination of feature extraction and kernel sparse representation, dictionary learning combined and kernel sparse representation improve the performance, experiments on ORL face database, Yale face database, AR face database, FERET face database and USPS handwritten database prove the value of the value of the algorithm in this paper.