Pose-varied Face Recognition Based On Multi-camera System

Two main reasons for affecting the performance of face recognition are face pose and illumination, and this paper proposes a fusion method of 2D face recognition and 3D face recognition in order to solve these two problems. In two-dimensional face recognition, face images are preprocessed to reduce the affect of illumination, and then Gabor features of faces are extracted, Fisher Linear Discriminative (FLD) is used to design classifier for two-dimensional face recognition. The basis of 3d face recognition is to obtain 3D face information. We use multiple cameras to get 3D information of face, and further 3D geometric features, with stereo vision technology, and then weighted minimum distance classifier is designed for 3D face recognition. Finally, the results of 2D recognition and 3D recognition are used for the final decision.For the research of 2D face recognition, we use Gabor features of faces and Fisher Linear Discriminant Analysis (FLD). The extraction of Gabor features is well studied, and Gabor features could extracted in different scales and different directions, and it could also reduce the influence of illumination by reducing the DC component in frequency. Because of multi-scale and multi-direction extraction of Gabor features, the quantity of data is greatly increased, and we use Fisher Linear Discriminant Analysis for reduction of data dimension.3D feature information which we access using multiple cameras is the basis of 3D face recognition, and we first get the images of faces with the same posture from different views, and then the position in three dimensional space of feature points is obtained using the technology of Stereo Vision.Thus, the 3D geometric features of faces can be determined. The three-dimensional geometry features we use include the distribution features of the key facial organ(distance between eyes, the distance from center of eyes to the tip of nose, etc), and include the shape features (mouth size, eyes size, etc). Finally, we use weighted minimum distance method for the classification and recognition. The main contents of face recognition are face detection, feature point positioning, feature extraction, face recognition, etc; We use stereo vision technology to access 3D face information, and the contents of stereo vision are camera model, camera calibration, three-dimensional matching, spatial positioning, etc. This research focuses on quick face detection, precise positioning of feature points, Gabor feature extraction and subspace technology, spatial point positioning based on the foots of perpendicular, and 3D face recognition based on 3D geometric characteristics.1) Face detection is the first step for face recognition. Our research is based on multi-camera, and this requires quickness of automatic face detection. Considering the images collected by the cameras are images with colors, the invariant features of face skin are used to detect the suspicious region including face. Adaboost algorithm is used to confirm if there is a face in the suspicious region, and this method could greatly reduce the areas to search, and improve the detection speed.2) The following step is locating of feature, and the precise positioning of feature points determines whether the face feature extraction is accurate and effective. Active Shape Model (ASM) is used for positioning of multiple feature points, and we improved the ASM algorithm by using 2D direction searching instead of one-dimensional direction searching for feature points. Although the computation is slightly increased, the convergence is quicker, which speeds the algorithm. The ASM algorithm can also positioning multiple feature points and the speed of positioning is fast, but positioning accuracy is not very high, especially with face pose greatly changed. So we use corner detection method to fix this problem.3) Finally, we design a system of face recognition with our algorithms. Illumination is the main influence factor of feature extraction in face recognition systems, so we spend much effort on image preprocessing. Homomorphic filter is to reduce the low frequency components and to increase the high frequency components with the purpose of minimizing the effect of illumination; The histogram matching method is used for consistency of the illumination on all the images. Then we build a new face recognition system based on two-dimensional features and 3D features, and this face recognition platform we established is a totally new system, and will be very helpful for our future research.