Research On Subspace Analysis Based Face Recognition Methods

Face recognition has been a focus in such fields as image processing and video precessing ect, its application has not become universal on real significance, relative to other recognition technology such as vehicle detection and recognition. However, in some specific fields of application, face recognition is still played a curcial role. Therefore, research on face recognition has attracted more and more researchs’attention and challenge.The critical issue of a successful face recognition technology is how to extract face features fast and efficiently. As one of face recognition research methods, subspace analysis method has become the mainstream for its simple computation, strong description and efficient identification. Based on the subspace analysis methods in application to face recognition, some relative research was made in this paper. The creative work of the thesis includes:1. Discriminant Method Based on Local Structure PreservingLocality Preserving Projections (LPP) is an algorithm of local feature extraction, it can preserve the local geometrical structure of the data set effectively. On the basis of LPP, Uncorrelated Discriminant Locality Preserving Projections (UDLPP) utilizes label information, and it achieves good discrimination capability by preserving the within-class geometric structure, while maximizing the between-class distance. A discriminant method based on local structure preserving called PCLSP is proposed with the idea of UDLPP, which both uses the label information of the data set and takes the local information of the data set into account. It enhances discrimination capability by minimizing the scatter of within-class neighborhood and maximizing the scatter of between-class neighborhood, and it will further use the local structure and increase the recognition rate. Experimental results on ORL and YALE face databases indicate that the proposed algorithm is effective.2. Dimensionality Reduction Algorithm Based on Pair-wise Constraints and Sparsity PreservingSemi-supervised learning is an algorithm of dimensional reduction by using some supervised information such as small number of labeled data or pair-wise constraints given by users, it has been widely used in the field of machine learning and data mining. This paper presents a dimensionality reduction algorithm based on pair-wise constraints and sparsity preserving. It combines some supervised information in the form of pair-wise constraints and large number of unsupervised information. It uses pair-wise constraints to discriminant analysis and uses sparse representation to preserve the sparse reconstructive structure in the transformed space. Compared with the traditional feature extraction method, this algorithm has a better recognition impact, lower parameters, and better robustness.3. Semi-supervised Dimensionality Reduction Based on Manifold Structure Preserving and Label PropagationLabel propagation algorithm is a graph-based semi-supervised learning approach. It propagates the labels from the labeled data to the unlabeled data iteratively by preserving the special structure of the whole data until a global state is achieved. Combined with label propagation algorithm and linear discriminant analysis, we proposed a semi-supervised dimensionality reduction based on manifold structure preserving and label propagation (SDRMPP). It propagates the labels by using the reconstruction weights as well as part of labeled data, the scatter matrices based on soft label learned by label propagation are constructed to perform the discriminant analysis. The feature extraction space was obtained by solving the optimization problem, and then the testing samples were classified into different classes. The experimental results on Yale and Feret face image database show that the proposed method is effective, especially when limited number of images were labeled, our method can still maintain its good classification performance.4. Marginal Discriminant Linear Local Tangent Space AlignmentMarginal discriminant linear local tangent space alignment (MDLLTSA) is proposed in order to solve the problems of LTSA in image recognition, such as implicitness of the nonlinear map or class information of data is ignored. With local tangent space representing for local geometrical structure of the manifold of the data samples, as well as with the supervised information for minimizing the margin of the intraclass and maximizing the margin of interclass. We convert the optimization problem of LTSA into multi-object optimization problem to obtain feature extraction space. Compared with classical feature extraction methods, the proposed algorithm obtains stronger classification as well as preserves local geometrical structure.