Face Recognition Based On Parameter-free Locality Preserving Projection

In the field of pattern recognition,face recognition has a broad development prospect.Although the in-depth study has obtained super achievements with security maintenance,intelligent monitoring and identity authentication,but many problems such as high-dimension dilemma and small sample problem have also arisen.This paper mainly aims at the feature extraction to solve them.Summarized as follows:(1)Summarize the historical background,the significance and applications of face recognition as well as several kind of common face feature extraction method.Analysis simply the feature extraction method and introduce several basic algorithms in the second chapter.(2)An improved minimum correlated locality preserving projections algorithm is proposed.To distinguish well different kinds of samples and solve the singular problem of matrices,the class information is introduced into locality preserving projections on account of weighted Maximum Margin Criterion.Second,neighbor graph is structured with cosine distance,which could settle parameter problem in LPP algorithm.Then,high dimension of face images can be reduced by singular value decomposition and redundant information is removed with minimum correlation analysis.Last,the experimental results on the ORL and Yale face databases validate the effectiveness of PCA?LPP?PCA + LDA and MMC.(3)The confluent MMC and 2D-DLPP algorithm is proposed.The 2D-LPP algorithm is introduced into the 2D-DLPP algorithm.Firstly,the 2D-DLPP algorithm is introduced into the MMC algorithm,which can maximize the distance between the sample classes at the same time minimum the distance within the class,as well as fully retain the local information and global information.Secondly,Pearson's correlation coefficient which describes the sample similarity can avoid the problem of parameter selection and redundant information can be removed by minimum correlation analysis.The contrastive experiments on ORL and Yale face database show that this algorithm has better recognition rate than 2D-LPP?2D-LDA?2D-DLPP.