Research On Face Recognition Based On Collaborative Representation And Linear Regression

Sparse representation based methods have shown excellent performance in face recognition. However, the later research shows that sparsity does not play a leading role in face recognition. Research suggests it is the mechanism called Collaborative Representation(CR) but not the sparsity constraint plays a critical role in sparse representation based methods. And the Sparse Representation based Classification(SRC) is regarded as a specialization of Collaborative Representation based Classification(CRC). This paper further illustrates it is actually linear regression plays a major role in these methods. This paper makes a review of some face recognition methods related to linear regression, and studies the linear regression in these methods. The main contribution and innovative points of the dissertation are summarized as follows:(1) The Linear Regression Classification(LRC) is analyzed. Aimed at the weakness that LRC is susceptible to the bias of sample, a two stage method called Extended Linear Regression Classification is proposed. The proposed method first divides the training images into different image sets according to environment factors like illumination, expression and pose. Then environment template images are created according to the former division. The template images are regarded as virtual samples and are used in the classification to improve the performance of LRC. Experimental results show the proposed method has a significant improvement in recognition rate compare to the original LRC.(2) We analysis the working mechanism of CRC, and illustrate the similarity in principal of classification between LRC and CRC. Considering that CRC is also affected by the bias of data samples, we propose a virtual sample based CRC. This method improves the recognition performance through expanding the dictionary using generated virtual samples.(3) In a viewpoint of linear regression, we make a detailed analysis of the relationship between CRC and SRC. Hence we extend the conception of CRC. At the same time, we propose a CRC based on elastic net, which calculates the regression coefficients using elastic net. Either CRC or SRC can be considered as a specialization of this method. And experiments show the recognition rate of this method is superior to the SRC and CRC.