Research On Methods Of Face Recognition Based On Manifold Studying Subspace

Face recognition is one hot topic in the field of pattern recognition, where feature extraction is the basic problem. In recent years, with the advantages of simplicity, high speediness and efficiency, manifold learning subspace algorithms are well implemented in feature extraction.The problems of manifold learning subspace algorithms are researched deeply in the background of face recognition such as extracting effective image features, removing redundant information among basis vectors, preserving structure information of image matrix, solving small sample size problem, solving nonlinear separated problems and so on.The main work is shown as follows:1. Traditional algorithms of face recognition are based on image gray matrix, while image gray matrix contains much more redundant information and is not sufficient for charactering the face feature information. In order to find out a feature representation method which is not influenced by face gesture, light, expression and so on, the Log-Gabor wavelet feature is extracted in face recognition. And then the modified unsupervised discriminant projection and local discriminant embedding algorithms are proposed. The method of removing redundant information among basis vectors is researched by adding extra constraints. And more over, the face recognition algorithm based on Log-Gabor and manifold learning subspace is given here, which takes advantage of both Log-Gabor and manifold learning subspace algorithms.2. Traditional manifold learning subspace algorithm always makes low dimensional projection on the eigenvector related to eignvalue of a ansymmetry eigen function. This can't guarantee the orthogonality among each projection vectors and will lead to information redundancy among feature vectors. The neighborhood preserving discriminant embedding is improved in this thesis. The orthogonal and irrelevant constraints are added in basis vectors. So an algorithm frame based on two basis vectors constraints is proposed. In addition, the algorithm is also extended to non-linear space which can deal with the problem of nonlinear problems.3. Original Isoprojection algorithm is an unsupervised subspace alogorithm, which can't utilize the regimentation information of samples efficiently. And it can not guarantee the space information among the pixel of an image which will generate singularity problem. In order to solve these problems, two direction supervised Isoprojection alogorithm based on image matrix is proposed. The class information of samples is introduced to improve recognition rate of Isoprojection. Also how to preserve the relationship between row and column direction of an image matrix is researched.4. Small sample size problem is often encountered in face recognition while traditional algorithms often lose valid discriminant information. In order to solve this problem, the constitution of feasible solution in small sample size is researched in marginal fisher analysis algorithm. In addition, the algorithms are also extended to non-linear space which can deal with the problem of nonlinear problems. The algorithm frame of getting optimal discriminant vectors in small sample size in kernel space is further researched.