Research Of Face Recognition Based On Symmetrical Subspace Analysis

Human face as a significant feature to human is important in emotion transmitting and identity recognition. Face Recognition research has great significance and extensive perspective among different kinds of biometrics recognitions because of its excellent characteristics such as non-touch collection, friendly, convenient and interactive. However, human face, which is a three dimensional non-rigid body, is easy to change because of facial expression and age_increasing. At the same time, face images that are used to recognition could be affected by the illumination and the environment. With the extended research, a variety of new algorithms have been applied to face recognition. But so far, no method is applicable to all cases, so it is still a great challenge to create a high-performance and robust face recognition algorithm in the research area. Feature extraction and feature selection are the key parts of the face recognition system; they are one of hot topics of face recognition area.This paper studies the linear subspace analysis based face recognition methods. Subspace analysis methods try to find a suitable linear or non-linear transform according to a specific performance target, and reserve the features, which are most useful to classification under statistic analysis, in the subspace. The distribution of the data is more compact in the subspace. Subspace analysis can reduce the dimension of the data, and present a more effective representation. Subspace analysis based face recognition methods utilize the statistic information of the face image; it has been one of the most popular methods because of low computation, strong representation capability and high separability. This paper improves the traditional linear subspace analysis methods and presents two new feature extraction algorithms, and applies them to three different face databases. Finally, tests their performance under different environment. The main contributions of this thesis are as follows:1. Presents a feature extraction method which combines the symmetrical characteristic of human face, odd-even decompose principle of function and weighted principal component analysis. First of all, this method makes use of the natural architecture characteristic of human face and transforms the face space to the odd symmetrical face space and even symmetrical face space, so that there is more information to face recognition. Then, applies weighted PCA to the odd-even symmetrical face space, makes the feature vectors of the symmetrical space equal with each other to reduce the effect of several principal components which involved by changes of illumination, facial expression and pose.2. Extracts the Independent Circular Symmetrical Gabor features (ICSGF), and applies them to face recognition. Circular symmetrical Gabor transform modifies the base functions of Gabor transform, and reserves the time-frequency or space-frequency locality. Then, applies the independent component analysis to the transformed data to realize dimension reduction and extracts ICSGF. The experiment results on ORL, Yale and FERET face databases demonstrates that this method is more effective than traditional Gabor based face recognition methods and subspace analysis based face recognition methods.