Face Recognition Based On Maximum Spacing Structure To Keep Projection And Sparse Representation

With the development of high-speed information technology era in21st century, it greatlyenriched the public access to various data sources, coupled with hardware development speed andthe speed of software development to keep pace, making the information we have obtainedshowing a high-quality, high-definition, large-capacity and high-dimensional features, which ledto "the curse of dimensionality." In this situation, manifold learning algorithms have emergedwith the sparse representation of the theory of algorithms designed to solve high-dimensionalnonlinear data dimensionality reduction for complex high-dimensional data. In recent years,many scholars of the above methods are widely applied to data mining, machine learning andother fields.In this paper, convection manifold learning and dimensionality reduction method based onsparse representation of the sparse representation classifier and manifold learning algorithmsin-depth research, and made a number of improvements, and we applied the proposed algorithmfor face recognition system design,we analysis and appraisal by a large number of experimentsshow that the algorithm is effective. Main results of the paper are as follows:1The paper analyzed and compared Some of the classical linear and nonlinear datadimensionality reduction methods （PCA, LDA, ICA, IOSMAP, LLE, LE, etc.）, it built a strongfoundation for the maximum margin structure preserved projection algorithm proposed.2We proposed maximum margin structure preserved projection algorithm（MMSPP）. Thismethod has the following characteristics（:1）in the face image sub-mode structure is divided, bythis mean,the nonlinear high dimensional data combine local and global structures;（2） Byintroducing LPP to remain space structure from the same face image of each sub-pattern;（3）By introducing NPE to remain different face images in the same position in the sub-patternstructure unchanged;（4） Supervised learning criterion MMC are introduced in this paper, sothat the algorithm can effectively utilize the training sample class information;（5） theorthogonal constraint are introduced in the objective function, to remove the correlation betweenthe feature vectors to improve the recognition rate.3In MMSPP algorithm,we proposed an improved method to replace the ordinaryneighborhood graph construction strategy to construct graph structure,it can solve "short road"and "isolated edge",experiments result shows that the improved constructed neighborhood graphis more reasonable.4An improved sparse representation classificatio（nSRC）is proposed in this paper. We usedordinary manifold learning algorithm and MMSPP to reduct dimensionality,at the same time,themethod combined the majority voting mechanism and SRC to design face recognition system, according to a large number of simulation experiments to verify the reasonableness of ourmethod.