Robust Image Recognition Algorithm Based On Sparse Constraints From Atom To Structure

Image recognition has being a research focus in the field of computer vision andpattern recognition. It has a wide range of application in military area, civil area andcommercial area. In recent years, the ideas of signal sparse representation have beenapplied to image recognition successfully and become the new concerns. How to constructa sparse representation recognition model, that is accurate, robust and fast, becomes anurgent question to study. In view of this, this paper put forward three new sparserepresentation recognition algorithms based on the analysis and summary of the relevantresearch.Firstly, considering the complementation of global and local information, bi-L1sparserepresentation algorithm for face image recognition based on the fusion of global andseparated components is proposed. First of all, the face image is used to obtain the globalsparse approximation based on L1sparse representation. Then several important facecomponents are extracted and aligned. After that, the sparse representation of all thecomponents is obtained respectively. And the sparse approximation results of eachcomponent are combined with a similarity voting method based on the approximationresiduals. Lastly, in decision-making layer, both of the sparse approximations are weightedintegrated to achieve classification. Experimental results verify the effectiveness of thealgorithm.Secondly, maximum likelihood sparse representation algorithm for image recognitionbased on class-related neighbors subspace is proposed. First of all, based on the differentdistribution characteristics of each test sample and the class-representative principle ofselecting training samples, local neighbors of adaptive number selected from each classare used to construct the new dictionary. Then the fidelity of sparse representation isrepresented by the maximum likelihood function of residuals and the maximum likelihoodsparse representation model is used to achieve robust classification. Experimental resultsdemonstrate the rationality of the algorithm.Finally, two modified structured sparse representation algorithm for image recognition is proposed aiming at the two problems that the selection of the sparsecriterion and the division of the blocks in the dictionary. First of all, the thought of serialcombination of the structured sparse criterion and the atom sparse criterion is proposed.The dictionary is reconstructed after structured sparse representation, and then the atomsparse representation is used to achieve classification. Then, structured sparserepresentation algorithm for recognition based on Maximal Linear Patch is proposed. Theimages in the same class are divided into blocks based on Maximal Linear Patch first.After that, the test image is recognized by structured sparse representation. Experimentalresults show that both of the algorithms are feasible.