Research On Face Recognition Based On Sparse Representation

Face recognition is an important technology in biometric identification, it is also the hot research topic in the field of image processing, computer vision, machine learning and so on, in the public security system, insurance, banking, customs, ID card system and other fields it has a broad application prospect.Different from the traditional Eigenface, Fisherface algorithm, In 2009,the face recognition based on Sparse Representation (Sparse Representation-based Classification, SRC) was proposed by Wright et al, because of its robustness to noise, it achieves a huge success in face recognition and it introduces a new development direction for face recognition. In this thesis, the sparse representation of face recognition is studied. The main work of this thesis are as follows:(1) Face recognition based on weighted sparse neighbor representation. First, in each class of the training samples, k samples are selected which nearest to the test samples to constructed a new training dictionaries of this class, then solving sparse coefficient with l1 norm minimization, a weight is given to the sparse coefficient of each new training sample; finally with the new training dictionary, according to the minimum reconstruction error to complete the recognition task. The most experiments results on extended Yale B face database and ORL face database show that the WSNRC method achieves higher recognition rate compared with NN and SNRC(Sparse neighbor representation for classification), confirmed the effectiveness of this algorithm.(2) Face recognition based on regularization fisher analysis and sparse representation. First, use the regularization Fisher analysis algorithm to extract the best projection matrix from the training sets, then we can get the low-dimensional representation of training sets and test sets under the projection matrix, finally, using sparse representation classifier for face recognition. Experiments results on AR database and extended Yale B databese show that the method of regularization fisher analysis and sparse representation is more effective.(3) Face recognition based on sparse representation with Gabor features and symmetric face. First the corresponding virtual symmetrical face can be obtained on the basis of original training samples, then combined the original training samples and their symmetrical faces to constitute the new training samples, finally extracted the Gabor features of all training samples and testing samples and use the SRC classifier for face recognition. Experimental results on the ORL database, Yale database and FERET database show that the effectiveness of GMSRC.