Manifold Learning And Tensor-based Multi-pose Face Recognition Research

In face recognition, if the images are not taken in front of the face, the face identification rate will be greatly reduced. If we can estimate the head pose then recognize the face with the appropriate view model, it will increase the non-frontal face recognition rate. The tensor face approach makes the facial image as a combined result of various factors: geometric structure, pose and illumination and so on, and it can decompose these factors. Manifold learning methods can arrange the view subspace into a continuous pose manifold, and estimate the approximate view of the new sample. Consider that when we research the multi-view face recognition by using the tensor methods, the first we estimate the view scope of the new non-identified face by manifold learning, and then we only project the new face to the identity subspace which in the calculated view scope to recognize. So, if the face poses has a greater change, it can get better recognition rate. According to this idea, the paper made the following work:1. Comparisions of the face view estimate methods which based on manifold learning:The face view estimation methods based on manifold learning: LLE(Locally Linear Embedding),LE(Laplacian Eigenmap),LEA(Locally Embedded Analysis),LPP(Locality Preserving Projections), the LEA,LPP are the nearly linear methods of LLE,LE. In order to choose a method with low estimate error as view estimate method in this paper. By comparing , we choose the LEA as the view estimate method.2. Improved the face view estimate method of LEA:Local embedding analysis (LEA) , in head pose estimation, selects local neighborhoods only from the same pose, the geometrical and topological information among the neighbor poses would lose. The paper presents an improved method of selecting neighborhoods to makes full use of the priori pose information. It makes the points of the same pose get closer while the points of different poses get further, so as to reduce the pose estimation error. The experimental results in Facepix database show the effectiveness of the proposed method.3. Proposed a method of muti-view face recognition based on pose estimation which combined the tensor method and manifold learning:The traditional tensorface method of recognizing face, there are the following disadvantages:(1) It projects the new face which would be recognized to the tensor identity subspace with all views, which increases the amount of computation and also reduces the recognition rate.(2) The tensor face is high dimension, when reduces the tensor to low dimension, the traditional method uses PCA, which is a linear method, but the tensorface is a high dimension with non-linear structure, dimension reduction with PCA would loose the non-linear information.This article has been improved on these defects: It first calculated the nearest view of the test sample in the training set with the Improved LEA of this paper, so it did not need to project the new face to all views, this will be high efficiency and recognition. We use LLE to reduce the tensorface instead of PCA, LLE is a non-linear method,when reduces to low dimension, it will hold the non-linear structrue in high dimension. This method has resolved the problem of lack of nonlinearity that exists in tensor, when the faces has large view changes, it also has high recognition. Also, experiments in this paper showed the effectiveness of the proposed method.