Sparse Representation For Face Recognition Based On Commonality-particularity Dictionary

Nowadays,face recognition technology has made great progress,but it faces many difficulties and challenges in practical applications.Many new ideas and algorithms have sprung up.Sparse representation has attracted the attention of many researchers because of its robustness to noise and occlusion and excellent classification performance.In particular,when Wright et al.have successfully applied the sparse representation to the face recognition field,face recognition enters a new climax.The idea of sparse representation for face recognition is to represent a test image by looking for the sparsest linear combination of the atoms in the dictionary,and assigns the test sample to the class which leads to the minimum reconstruction error.Therefore,the discriminant ability of the dictionary has an important influence on the classification performance.Various dictionary learning algorithms is proposed and improved continuously.Takes into account the commonalities of different categories that do not contribute to the classification,the commonality-particularity dictionary,a classification oriented discriminative structured dictionary is designed on the basis of preserving the particularities of the samples.The commonalities of the dictionary mainly represents and reconstructs the data,and the particularities mainly distinguishes the data.The good discriminability advantage of the commonality-particularity dictionary is of great help to the improvement of the accuracy of face recognition.This thesis studies sparse representation for face recognition based on commonality-particularity dictionary from two aspects:the discriminability of dictionary and the discriminability of sparse representation,the corresponding algorithms are modified,and the effectiveness of the proposed algorithms are verified on open databases.The main work and innovations are summarized as follows:(1)Locality-constrained sparse representation for face recognition based on commonality-particularity dictionary is proposed.In order to excavate the implied local structure information in the training data based on the good discriminability advantage of the commonality-particularity dictionary,the improved local constraints to the sparse coding process of test samples is applied,the similarities between the test sample and its adjacent atoms of particularities is retained,thus a more discriminative sparse representation coefficient is obtained.In addition,two methods of similarity measure are given in view of the special structure of the commonality-particularity dictionary.Several experiments on the Extended Yale B database,the AR database,and the ORL database verify the effectiveness of the proposed algorithm.(2)Weighted group sparse representation for face recognition based on commonality-particularity dictionary is proposed.It had been verified that more discriminative representation can be obtained by group sparse representation compared to the traditional sparse representation.In order to use as few as possible dictionary groups which is close to the test sample to sparse represent the test sample,the test samples is weighted group sparse represented based on commonality-particularity dictionary,First,we use the reconstructed error values of test samples on all categories of particularities and commonalities to predict the category of test samples.The prediction value is then used as the weight of the sparse representation coefficients of each group in the weighted group sparse representation.The face recognition algorithm makes full use of the discriminant performance of the commonality-particularity dictionary and the representation ability of the weighted group sparse,and the recognition performance of the algorithm is improved.Several experiments on the Extended Yale B database and the AR database verify the effectiveness of the proposed algorithm.