Research On Recognition Technique Of Face Manifold Based On Feature Description

Face recognition has become a hot research topic in the field of biometric identification technology for the merits of being directly perceived and intuitive. However, human face images always include different expressions and complex environment factors, it's crucial to extract effectively discriminant features for face recognition. Manifold learning has been proven to find the intrinsic feature information and the inherent laws in face data effectively. With variables(light, pose)changing, the sample data can be hypothesized to be on a smooth surface in high-dimensional space, so face images can be regarded as a low-dimensional manifold embedded in the high-dimensional space. However, manifold learning applied in face recognition is usually based on whole image vectorized description, which makes recognition algorithm suffer from noise. Meanwhile, a large number of studies indicate that a much more stable description can be obtained by using local features, which can improve the recognition rate. In this thesis, local features are used to improve existing face manifold algorithm to improve the robustness of the system.This thesis analyses face manifold recognition algorithm and propose a face manifold model based on feature description to optimize it. This can be achieved by two schemes. The first scheme applies Local Binary Pattern to the whole face image, then a discrete cosine transform is used to obtain the face manifold description in the transform domain. Experiments show that this program can achieve a high recognition rate in the benchmark datasets, and can eliminate light impacts effectively, however, it fails in unconstrained face datasets. The second scheme uses facial landmark methods to locate a set of ordered points, and SIFT features are computed. These feature vectors are merged by these points' order, which is named a face manifold based on local feature description. Facial landmark technique can ensure the sparsity of sample features and stability of face manifold. What's more, plenty of experiments have been conducted to compare ASM(Active Shape Model) and ESR(Explicit Shape Regression), and the latter is proven to be better. Experiments show that this method can obtain a relatively high recognition rate in unconstrained face dataset, and the crucial part of this scheme is how to select description area and the distribution of facial landmarks. At the end of this thesis, an online face recognition system is presented by the second scheme.In conclusion, face manifold based on feature description is more robust than whole image vectorized description based method. Meanwhile, SIFT feature can adapt to varying expression and pose, as well as occlusion and disguise. In brief, compared with the global transformation domain features, such as LBP etc., SIFT feature has higher capacity to describe face images.