Manifold Learning And Sparsity Preserving Projections Based Face Recognition

Face recognition is a research focus in pattern recognition; it is widely used in variousapplications such as remote intelligent monitoring， human-computer interaction and securityprotection. Recently, sparse representation has been used in face recognition, and achieves goodresult. Sparsity Preserving Projections (SPP) is an effective face recognition method which isdeveloped on the basis of sparse representation. SPP finds a low-dimensional space which canpreserve global reconstruction relations among samples. However, in practical applications, imagesamples reside on a nonlinear manifold of the high-dimensional space, which is the inherentstructure among the samples. In this paper, we combine SPP with manifold learning, and propose aseries of novel methods.Firstly, we propose Manifold Learning based Uncorrelated Sparsity Preserving Projections(MLUSPP) approach. The basic idea is that we combine SPP with manifold method. As a result, theproposed method can not only preserve global reconstruction relation but also preserve manifoldstructure among samples. Furthermore, we use uncorrelated constraints to remove redundantinformation in discriminant features.Secondly, MLUSPP is an unsupervised method while supervised method can use classinformation to improve recognition performance. We combine MLUSPP with supervised learning,and propose Manifold Learning based Kernel Uncorrelated Sparsity preserving DiscriminantAnalysis (MLUSDA). MLUSDA uses class information to improve similarity graph, which canpreserve manifold structure better.Thirdly, in order to further enhance the dicriminability of MLUSPP and MLUSDA, we extendthem to the nonlinear space, and propose Manifold Learning based Kernel Uncorrelated SparsityPreserving Projections (MLKUSPP) and Manifold Learning based Kernel Uncorrelated Sparsitypreserving Discriminant Analysis (MLKUSDA).Experimental results on FERET, AR and CAS-PEAL databases demonstrate that the proposedapproach can effectively improve the recognition results as compared with some related methods.