| Face recognition is to recognize or validate one or more faces in static or dynamic scenes, using technology of image processing and pattern recognition. In most practical applications of face recognition, the imaging condition is uncontrolled. It is impossible to keep the illumination invariant and to require the cooperation of human object, therefore, face recognition with illumination or pose variation is one of recent study focuses. In this thesis, a framework of face recognition system robust to pose variation is proposed. The details are as follows:1. An improved Cascade MR-ASM(Multi-resolution Active Shape Model) method is proposed, which combine cascade with minimum policy to locate landmark point exactly, based on Euclidean distance of face texture.2. The face alignment methods are studied in both 2D and 3D. The location results are mapped into the public parameter space and the shape parameters are obtained. Then the frontal ASM model is obtained by setting zero to parameters related to pose. The aligned face is obtained by texture mapping. One the other hand, 3D model of human face is recovered from face model of two face images. The frontal face model is obtained by linearity transformation in 3 dimensional spaces. Finally, an aligned face image is obtained by texture mapping.3. Facial feature extraction and face recognition are studied. We extract two kinds of facial features for face recognition. The one is the features using eigenfaces. The other is Local Gabor Texture (LGT), which uses LGT histogram for face recognition. The face recognition is implemented using K-near Neighbor Classifier. The experimental results using two features are compared.4. A face recognition system is designed and implemented. The system is designed by one of the international standard in biometrics BioAPI. It involves face detection, location of landmark points and face recognition. The system can process both still face images and the face in surveillance videos. |
PDF Full Text Request