Robust Image Recognition Algorithm Of Sparse Representation Based On Sample Extended

Image recognition as an important part of the information science and signalprocessing field, is widely used in information security, military, production and living, etc.In recent years, the ideas of sparse representation have been applied to image recognitionsuccessfully. Because of its high recognition rate and strong robustness etc, sparserepresentation for image recognition has become research hot spot. This paper will raisethree new sparse representation algorithms for image recognition based on the analysisand summary of the domestic and international relevant research results.Firstly, because the data structure information plays a certain support role, we applythe block-structured sparse representation for face recognition. Considering the problemthat the training images can not span the facial variation under testing conditions, a novelrecognition method of joint bi-sparse representation based sample extended differencetemplate is proposed. It applies an extended difference template to represent the possiblevariation between the training and testing images. Because the intra-category variation ofany category can be represented as the atomic sparse linear combination, the recognitionproblem is converted into finding a joint bi-sparse representation of the block-structuredsparse representation and atomic sparse representation in the training sample space andextended difference template space for classification.Secondly, in consideration of the cast shadows, specularities, occlusions andcorruptions in the images that violate the low-rank structure, we propose a novelrecognition method of joint sparse representation based on low-rank subspace recovery.First of all, we can decompose the data matrix which is composed of all training images ofeach class as the sum of a low-rank matrix and a sparse error matrix, this two partsrepresent the ‘clean’ image which follow strictly the low-rank subspace structure and theerror which violate the low-rank structure respectively. Then the test sample can berepresented as the linear combination of dictionary which is composed of low rank matrixand error matrix, using the sparse approximation of this two parts calculates the residualwhich used for classification. Finally, sub-modular sparse representation algorithm for image recognition based onBorda voted weighting is proposed aiming at the problem of the discriminative criterion ofsub-modular sparse representation and of that each sub-module plays different roles in thefinal classification. When exploiting sub-modular sparse representation for classification,the Borda method is used for count where each class is assigned to different votesaccording to the residual size. Then it combines the sub-modular sparsity and sub-modularresidual jointly to determine the sub-modular credibility weight. After that, according tothe sub-modular credibility weight, we calculate the votes of each class of all thesub-modules for the final classification.