Research On Face Recognition Based On Graph-optimized Two-dimensional Locality Preserving Projections

Face recognition, as one of the import branch of pattern recognition, has importantscientific and practical value, and it also has been attached great importance to the researchinstitutions and enterprises. Many face recognition methods, which is based on LPP(locality preserving projections) algorithm has received wide attention from researchers inrecent years.The LPP algorithm has two advantages:the local structure of images space canbe preserved; it is a linear method. compared with the general linear dimensionalityreduction algorithm, the algorithm of LPP has manifold learning ability. When the highdimensional data hidden in the low-dimensional manifold structure, LPP approximatelyobtains the best projection direction by usingthe optimal linear approximations to theeigenfunctions of the Laplace Beltrami operator on the face manifold. Because theadjacency graph of the LPP algorithm is artificially defined in advance, without consideringthe needs of the subsequent reduction, and constructing adjacency graph depends onexperience, it is difficult to determine the number of nearest neighbors; Gauss function fordetermining nearest neighbor graph weights is also difficult to determine. In order to solvethese problem,L. Zhang proposed a novel algorithm called Graph-optimized LocalityPreserving Projections (GoLPP). This paper focuses on the GoLPP algorithm, and as wellas applys the graph-optimized theory to2DLPP algorithm, and then I proposes Go2DLPP(Graph-optimized two-dimensional locality preserving projections) algorithm.The main work of this paper includes:1) researching on the face recognition algorithm based on LPP (LPP,2DLPP), as thetraditional LPP algorithm based on adjacency graph does not consider class labelinformation of image samples, the paper adds the class label information into theconstruction of the adjacency graph, Experimental results show the efficiency of thealgorithm is improved. 2)researching on the GoLPP algorithm.The idea of the GoLPP algorithm is to integrategraph construction with specific dimensionality reduction process into a unified framework,which solves the problem of parameter selection occured in LPP algorithm,and alsoimproves the algorithm efficiency.3) applying the graph-optimized theory to2DLPP algorithm, and then develop a novelalgorithm called Graph-optimized Two-dimensional Locality Preserving Projections(Go2DLPP). Doing experiments on the Go2DLPP algorithm and GoLPP.Experimentalresults show that the efficiency of Go2DLPP is better than GoLPP and the traditional LPPalgorithms.