Research On Face Feature Extraction Algorithms Based On Linear And Nonlinear

Feature extraction is the most crucial part of face recognition system, whichhas important role in the field of pattern recognition. There are a lot of classicalgorithms about feature extraction at present, such as the method based on linearand nonlinear. The subspace method has been widely used with therepresentative methods, including the principal component analysis (PCA), lineardiscrimination analysis (LDA) and locality preserving projection (LPP). LPP islinear approximation of the Laplacian Eigenmap (LE), and there emerged manyeffective methods on the basis of the improvement of LPP. However, the abovemethods are limited to dealing with the linear inseparable problem of face images.Therefore, this paper introduced some nonlinear feature extraction methods onthe basis of the linear feature extraction method, such as the kernel technologyand kernel extension of graph embedding algorithmThis paper includes the following three parts:1. This paper proposed the method of the orthogonal complete discriminantlocality preserving projections of kernel (KOCDLPP). After combining theconvergence kernel technology and orthogonal theory, the objective function ofdiscriminant locality preserving projections got changed from business operationto difference operation.2. This article studied the discriminant locality preserving projectionsmethod, which existing the small sample problem, and come up with thediscriminant locality preserving projections method based on neighborhoodmaximum margin (NMMDLPP). Firstly, the weighted K-nearest neighbor graphof training sample was constructed, and the local geometry information of theintraclass neighbors and interclass neighbors of each point was obtained bygiving weight to each side of the nearest neighbor graph, then there also got transfer matrix by reducing the interval between the intraclass neighbors andincrease the interval between the interclass neighbors, the neighbor point optimalrefactoring coefficient of the data was applied to the objective function. Thismethod took the difference between the locality preserving between-class scatterand the locality preserving within-class scatter as the objective function in orderto avoid the calculation of the inversion of matrix.3. The Kernel Extension of Graph Embedding Based on CanonicalComponent Analysis (KGE/CCA) is proposed based on the CanonicalComponent Analysis (CCA) algorithm and Kernel Extension of GraphEmbedding (KGE) algorithm. This method not only parted the feature extractionof sample classification information and nonlinear, but also maximize thecorrelation by combined the two features into one feature, which increased theamount of information with uncorrelated statistics.The effectiveness of the three algorithms proposed by this paper is presentby a large amount of experiences in the standard face databases, such as UMIST,JAFFE, Yale, etc.