Face Recognition Based On Subspace Feature Extraction

Feature extraction is an essential question in Face Recognition, which is the key toimprove the recognition performance. Feature dimension reduction methods include PrincipalComponent Analysis (PCA), Linear Discriminant Analysis (LDA), Local PreservingProjection (LPP), Sparse Representation based methods and so on. In this paper, we mainlystudy subspace based feature extraction algorithms, namely LDA and Sparse Representationbased method. Considering the local structure of the samples, new subspace based featureextraction algorithms are studied, which are applied in Face recognition. The main works inthis paper are as follows:(1) This paper proposes Local Nonnegative Sparse Preserving Projection method(LNSPP). From the total sparse coefficients of the target sample, we found that most ofnon-zero coefficients belong to the neighbor of the target simple. In order to obtain moresparse coefficients and make better express the relationship between data, we use radial basickernel function to calculate the neighbors of the target sample, and then exploit theseneighbors to nonnegative sparse representation, finally look for a projection space thatmaintains the relationship of local nonnegative sparse reconstruction between all trainingsamples. The experiment results show that LNSPP can obtain better recognition than SparePreserving Projection (SPP).(2) This paper proposes Collaborative Preserving Projection method (LCPP). Thismethod incorporates collaborative representation into SPP, and uses the neighbors of thetarget training sample to collaboratively represent it, which can greatly reduce the runningtime of the LCPP method. And meanwhile, the recognition rate of LCPP has certainenhancement.(3) Fuzzy Local Linear Discriminant Analysis (FLLDA) is presented in this paper. In theresearch of Local Linear Discriminant Analysis (LLDA),we found that it uses Euclideandistance to seek neighbor samples, and the distribution information of neighbor simples is notconsidered. With the view of this point, FLLDA uses kernel function to obtain the neighborsamples of a given testing sample, and then uses the fuzzy memberships of the neighbors toredefine the within-class scatter matrix and the between-class scatter matrix. The method canmake full use of the distribution information of the neighbors. The experiments results onORL, AR, FERERT face databases show that the effectiveness of the proposed method.(4) This paper proposes Collaborative Local Linear discriminant analysis (CLLDA). In CLLDA, we use collaborative representation to calculate the distances between the testingsample and training samples, and then obtain the neighbors of a given testing sampleaccording the distances; finally we use the neighbors to linear discriminant analysis. Theexperiments results show that the effectiveness of the proposed method especially on AR,FERET face databases, and the proposed method is more stable than LLDA algorithm.