Subspace Methods For Face Recognition Research And Realization

Face recognition (FR) has become an increasingly popular subject in the field of biometrics. More and more researchers in Computer Vision and Pattern Recognition pay their attention to this field. Feature extraction is one of the most important subjects to achieve high recognition performance.Subspace method is a popular featuter extraction method in the FR task. Subspace method wants to find out a conversion with the ability of representing the data set with the effective features in minor dimension space without decreasing the inherent information contained in original data. In this thesis several popular subspace methods have been analysed and successfully realized in MutiBIS FR sub-system.However, each kind of subspace method might loss some important features. So a FR algorithm is designed by fusing holistic and local features in subspaces, which is inspired by Human Visual System (HVS). The holistic features are extracted by principal component analysis (PCA), and the local features are extracted by non-negative matrix factorization with sparseness constraints (NMFs). Liner discriminant analysis (LDA) is applied to enhance adaptive ability to illumination and expression. Fusion algorithms are designed at feature extraction level and matching score level, by concatenating feature vectors, and integrating the matching scores using an Adaptive Neuro-Fuzzy InferenceSystem (ANFIS), respectively. The experiments on UMIST face database show that fusion scheme outperforms individual algorithm based on PCA or NMFs.This thesis is organized as follows. In chapter one, the basic conception of face recognition, its application and research fields are introduced. The subspace method and several important algorithms are described in chapter two. In chaper three, the fusion scheme is presented. The development of MutiBIS FR sub-system is discussed in the next chapter. Conclusions are drawn in the last chapter.