Methods For Facial Features Localization And Face Recognition

The research work of this paper focuses on the localization of facial features and face recognition. In our researches for facial features localization, active shape model (ASM) and active appearance model (AAM) are used, and a new method is put forward based on the combination of local texture model and global texture model. Improvements are made to the standard ASM for facial feature localization. Local area constraint method and edge constraint method are proposed. Also, a points labeling strategy which increase the relativity of the labeled points is put forward, and a semi-automatic feature points labeling tool is designed which greatly improve both the accuracy and the efficiency of labeling work. Experiment shows the improved active shape model performs significantly better than standard active shape model in the localization of facial features. Active appearance model (AAM) is also studied for the localization of facial features. The parameters of texture model and appearance model are studied in detail in reconstruction of face images. Based on the analysis of the theory and capability of ASM&AAM, a new method is put forward by combination of local texture model and global texture model. Not only the local texture model in ASM is adopted in this method, the global texture is used as well to restrict the result and to predict a new shape. In order to calculate the matching degree of a shape to target image, shape free texture subspace reconstruction error was adopted as matching criterion. The reasonability of this method is supported by our test. With this criterion, searching method based on local texture model and the one base on global texture prediction are connected. In the whole progress, the matching degree of the result shape keeps rising. Hence, the robustness of this algorithm is raised.A face-recognition algorithm based on Fisher linear discriminant analysis is studied in detail which combines principal component analysis (PCA) based Eigenface method and linear discriminant analysis (LDA) method. In the process of extracting face features for recognition, pure PCA method can obtain the best features in the sense of representation, Butthese features are not very suitable for classification purposes. LDA method use label information to get the best features for classification purposes. But the dimensionality of the original face image is too high to use LDA method directly. The algorithm studied in this paper consists of two steps: first reduce the dimensionality of the face image via PCA, and then use LDA to obtain the best projection direction for classification purposes. The improved performance of this new algorithm over pure PCA algorithm is demonstrated by our experiments. Based on this algorithm, an automatic face recognition system is put forward which integrates face detecting, iris location, and face recognition algorithm. A robust real-time face recognition result can be achieved by analysis the consequent video images.