Research, Facial Image Recognition Method Based On Sparse Representation

Face Recognition is considered as an important technology of modern biological information recognition. For a given face image, confirm the identity of the image by using the stored face image database. Most of the exsiting face recognition methods need complex image preprocessing and feature extraction, the choice of features has a great effect on the recognition rate, and it lacks robust about occlusion and corruption. These problems often make the existing recognition methods restricted in the application.With advantages like high recognition rate and strong robustness, more and more researchers take attention on sparse represention. Sparse represention is the important theory of compressive sensing. Sparse represention of data can reduce the cost of data processing substantially, increase the efficiency of compression. Sparse Represention has a unique advantage in classification, it makes feature selection is no longer critical. The work of this paper are as follows:(1) Research the theory of compressive sensing and sparse represention, and it indicates that sparse represention has natural discrimination, it chooses the most compact subsets of the signal representiont. Thus, sparse represention can be used for classification.(2) Use the discrimination of sparse represention for face recognition. Represent the test sample as the linear combination of training data. Propose a face recognition algorithm based on sparse represention, combined with features such as downsample, Eigenface, Laplaction, Fisherface, Randomface and so on. Then present experiments on the Extended Yale B and ORL face databases, and the results indicates proposed algorithm obtains higher recognition than traditional methods.(3) Improve the robustness of the proposed recognition algorithm. Add the validity discrimination, and add the error part into the original model, make the algorithm can deal with the occlusion and corruption. Considering the low recognition rate of the unmatched face image, we proposed a automatically match and recognition algorithm.(4) Use the great ability of total variation on describing the detail, we propose to obtain the sparse solution by total variation instead of the minimization l1-norm. It makes up the accuracy problem due to the definition of the minimization l1-norm. Then we can obtain the more sparse solution and higher recognition rate.