Research On Face Recognition Based On Subspace Methods

There is a lot of redundant information in primitive face images, how to discard theredundant information and obtain the most discriminative features in order to identify thetarget efficiently, accurately and in real time, which is the key to the research on facerecognition. In this paper, how to extract discriminate information effectively from theraw face images is our research objective, which focuses on subspace methods on thefield of face recognition. Although subspace methods can reveal the potential lowdimensional sub-manifold structure of high dimensional data very effectively, they stillhave many deficiencies, so scholars have proposed lots of algorithms to improve themfrom different aspects. In this paper, the original algorithms and their variants are studied,further exploring their potential problems and proposing the corresponding solutions,mainly including the following two aspects:(1) Neighborhood preserving embedding(NPE) directly reconstructs the sample byits k-nearest neighbors, however it does not distinguish between the importance ofintra-class neighbors and inter-class neighbors, leading to poor recognition performance.In this paper, a fuzzy neighborhood preserving embedding algorithm based on commonvector(FNPE/CV) is proposed. Firstly, the degree of membership of every sample in eachclass is obtained based upon the class label of its k-nearest neighbors. And then everysample is reconstructed by the common vector and its membership grades for every class.Finally, the problem of minimizing the residual between the original sample and itsreconstruction sample is converted to solve the generalized eigenvalue problem in orderto obtain the final projection transformation matrix. After the projecting, FNPE/CVminimizes the difference among intra-class samples and separates inter-class samples assoon as possible. Experiments on four face databases(such as ORL, Yale, AR andPIEC29) demonstrate the effectiveness of FNPE/CV.(2) Locality preserving projections(LPP), as well as neighborhood preservingembedding, which could preserve the neighborhood structure of the data set efficiently, however, too much emphasis on the local geometric feature always leads to globalgeometric feature may not be kept well. Since LPP and NPE are unsupervised, which isthe other major defect of them, the classification performance of them is not so well. Forthese drawbacks, a discriminant preserving embedding(DPE) algorithm based oncorrelation coefficient is proposed in this paper. A new and concise discriminant featureextraction criterion is raised by minimizing the intra-class similarity scatter, minimizingthe modified intra-class reconstruction residual and simultaneously maximizing theinter-class reconstruction error, which could preserve both the local geometric propertiesof the data and the overall geometrical structures effectively. The results of extensiveexperiments on four public face databases(such as ORL, Yale, YaleB and PIE) illustratethe effectiveness of DPE.