Discriminant Neighborhood Structuer Embedding And Image Recognition

Image recognition is a basic research field of pattern recognition. With thedevelopment of information technology, the dimension of the image is increasingly high.How to effectively describe and classify the image is the difficulty of the current study.Feature extraction can identify features conducive to subsequent analysis by a certainlinear or nonlinear transformation. We can explore the intrinsic geometry of thelow-dimensional representation of the data and reduce the computational complexity byfeature extraction. Therefore, feature extraction has become a hot topic in the study ofimage recognition technology. The main work is as follows:First, we propose Discriminant Neighborhood Structure Embedding (DNSE) tosolve the over-fitting and local geometry destroied problem existing in lineardiscriminant analysis and neighborhood. DNSE keep the global structure of the data byusing the advantage of linear discriminant analysis. DNSE keep the local geometry ofthe data by local similarity and local diversity adjacency graph embedding. Theover-fitting problems have been effectively overcome by local diversity adjacencygraph embedding. Experimental results demonstrate the performance of the algorithm.Second, studies show that low-dimensional description obtained by the orthogonalprojection vector is more conducive to classification. Trace Ratio solution is the globaloptimal solution of the objective function, and the solution satisfies all directionsorthogonal. Therefore, we propose Trace Ratio Based Discriminant NeighborhoodStructure Embedding (TR-DNSE). Compared DNSE, TR-DNSE use a set of orthogonalvector to the projection, and is more conducive to the subsequent identification andclassification. Algorithm is more robust. Experimental results demonstrate theperformance of the algorithm.