Subspace Algorithm For Face Recognition

As an important research direction in the field of computer vision,face recognition has been widely used in production and life,bringing great convenience to people's production and life.However,face images are often affected by a variety of conditions such as the orientation of face,occlusion,and expression,which brings challenges to face recognition tasks.Machine learning algorithms easily suffer from the curse of dimensionality in processing high-dimensional images.Meanwhile,it is found that the face images have an implicit manifold structure in the original space.Subspace algorithm represented by manifold learning reduces the dimensionality of high-dimensional images to extract its intrinsic low-dimensional features,avoids the curse of dimensionality,and retains the manifold structure of data,which has been widely applied in the field of face recognition.The global subspace algorithm represented by linear discriminant analysis maximizes the distance of inter-class samples and minimizes the distance of intra-class samples by using the global information of samples,which achieves better discriminant performance.While the global subspace algorithm only uses the global information and ignores the local neighbor structure of samples,so that the subspace cannot reflect the intrinsic features of data.Meanwhile,it is also suffering from the problem of small sample size(SSS).The subspace algorithm represented by the manifold learning algorithm usually keeps the local neighbor structure of samples in the original space and extracts the intrinsic features of the samples.However,it is necessary to use the global information of samples to enhance the discriminant ability of algorithms for the classification tasks.In the face recognition task,we should use global and local information at the same time to enhance discriminant performance and maintain the manifold structure of data.How to balance local and global information is a valuable problem.To solve the challenges in face recognition and subspace algorithms mentioned above,the researches of this paper are as follows:Firstly,Margin Discriminant Projection(MDP)algorithm is proposed for the weaknesses of the maximum margin criterion algorithm(MMC)and the margin fisher analysis algorithm.The intra-class scatter is defined by the class samples'mean and it's marginal intra-class samples,the inter-class scatter is defined by the class samples' mean and it's marginal inter-class samples.Meanwhile,the MMC is used to solve the singular value of the intra-class scatter matrix.MDP provides a general linear dimensionality reduction framework by using global information and the local information.Secondly,we propose the Locality Sensitive discrimination Projection(LSDP)algorithm to solve the intra-class outlier problem and maintain the inter-class neighbor relationship.We enhance the compactness of intra-class samples to solve the problem of intra-class outliers and maintain the local structure of inter-class neighbor samples to preserve their relationship in subspace.At the same time,the difference between inter-class scatter and intra-class scatter is minimized to avoid the SSS problem.LSDP is more robust to intra-class outliers and preserves samples' intrinsic features in subspace.Thirdly,in order to maintain the similarity information of intra-class samples and the diversity information of inter-class samples,Similarity and Diversity Discriminant Projection(SDDP)algorithm are proposed.Based on prior knowledge we introduce the similarity weights between non-neighbor intra-class samples and the diversity weights between non-neighbor inter-class samples.The introduction of non-local information makes better use of the potential pattern information between non-neighbor samples according to prior knowledge.At last,the MMC is used to avoid the SSS problem.