Research On Various Poses Face Image Registration

Most of the existing face recognition algorithms are based on frontal or quasi-frontal face images. But the face recognition rate decrease tremendously, when these algorithms are applied directly to non-frontal face images.Therefore, the pose of non-frontal image become a very important factor to affect the face recognition algorithms. In order to improve non-frontal face images recognition rate, a new hierarchical various poses face registration algorithm is proposed, which utilizes modified feature location algorithm and hierarchical image registration algorithm, to transform various poses images to a fused frontal image which contains richer information.The contributions of the thesis are as follows:1) A hierarchical face registration algorithm is proposed. The registration algorithm includes the modified feature location algorithm based on the Constrained Local Model, the coarse face registration algorithm based on priori knowledge of face and the fine face registration algorithm based on low-rank decomposition.2) A modified feature location algorithm based on the Constrained Local Model is proposed. After studying many face models, Constrained Local Model is used to construct face model, which combines shape and local texture. The preciseness of feature location is an important factor affecting the performance of the face registration algorithm. The thesis utilize the Reverse Combination Algorithm and Constrained Local Model to locate features. The experimental results show that the modified feature location algorithm is more accurate and efficient.3) A pose estimation and coarse face registration algorithm is proposed. The poses of non-frontal face sequence are different, most close to the frontal face image is chosen as reference image by pose estimation in coarse registration algorithm. As the location of features have already acquired, the poses of the non-frontal sequence are estimated, and the non-frontal faces can be transformed to reference image using the method of polynomial fitting. The experimental results show that a certain degree of correction have been acquired, further more, the image distortion is smaller.4) A fine face registration algorithm based on low-rank decomposition is proposed. After coarse registration, the non-frontal face sequence are more linearly correlated. In this case, low-rank decomposition is used in the dissertation to calculate transform matrix and remove sparse error. Utilizing the transform matrix, the non-frontal face sequence can be transform to the frontal pose.The experimental results show that the proposed hierarchical face registration algorithm has better performance among frontal face image synthesis methods.