Face Recognition Based On Subspace Of Circularly Symmetrical Gabor Transforms

The visual information reflected by a face plays a major role in conveying identity and emotion. Because of its extensive and applied realm, the research of face recognition has very important practical value. Face recognition has become one of the key technologies in biometrics and got the extensive concern with study. However, it is difficult to implement face recognition using computers. On one hand, human face is a deformable object composed of complex 3D curve surfaces, which is hard to be represented in the form of mathematics. On the other hand, faces of different persons have the similar structure and the face images are greatly dependent on ages, illumination and imaging environment.This thesis mainly studies face feature extraction which is the key part of face recognition. First, transform the face image with a certain transformation, and then extract the discriminant features for face recognition using subspaces methods. Following this feature extraction idea, we summarize the main work of this thesis as follows:(1) Circularly Symmetric Gabor Transform (CSGT)First, research is conducted on Gabor Transform (GT) based face recognition. Two typical approaches using Gabor features, Elastic Graph Matching (EGM) and Gabor feature Fisher Classification (GFC) are analyzed. Then we introduce CSGT with complete analysis and discussion to the definition, concept and properties of it. By comparing to the traditional GT, it is found that the former is remarkably superior in rotation invariance and reduction of data redundancy.(2) Weighted PCA (WPCA) and Modified Maximum Margin Criterion (MMMC)In the study of the traditional PCA approach, we found the different effects of different features in face recognition, and proposed to perform face recognition in a weighted PCA space. It is pointed out that classification by Euclidian distances in WPCA space is equivalent to that by Mahalanobis distances in the traditional PCA space. Thus, the reason behind the performance improvement of the WPCA approach is given in theory.Based Linear Discriminant Analysis (LDA) and Maximum Margin Criterion (MMC) methods, we propose MMMC method that redefines the within-class scatter matrix and the between-class scatter matrix. Experiment results prove that the new method is stable and effective.(3) CSGT-subspace method based face recognitionWe carry out a complete study to the problem of applying CSGT to face recognition and decide to use the magnitude feature of CSGT. First, down-sample the transform domain and form augmented feature vectors, and then extract classification features using subspace method. Thus we propose two algorithms: CSGT-WPCA and CSGT-MMMC. Experiment results on the ORL, Yale, and Feret face databases are shown and discussed.