| Based on the high identification of 2D facial texture features and the good robustness against variation of position, illumination, and expression of 3D face data, we focus on the geometric feature extraction of 3D face point cloud, feature extraction of 3D depth image and 2D face image and fusion methods of face recognition. Our main work includes:(1) While considering the particularity of the point cloud of nose region, we presents a method for registration using the local point cloud of face to correct face posture. Firstly, the registration performance and efficiency of the ICP algorithm are weighted to determine the area size. With the symmetry of face, the holes generated during data acquisition process can be filled by mirror. Experiment results show that it’s effective to correct posture by the area around the nose tip when we weigh the registration error against the costs of time.(2) About the point cloud geometric features, the most two representative contour lines of face are extracted. Considering the robustness against position and expression, nose area contours features without mouth and face edge are extracted. Then identify the local contour lines that extracted, and analyze the impact of different external factors. It shows the high discrimination of effective local contour features. Further more, recognition rate has been improved on different test set when fusing two contour features with different weights. The two contours had different influence on recognition rate.(3) Multiple information fusion is studied in detail. On the one hand, the weighted LBP feature of 2D face images and 3D depth images are extracted. They are weighted to fuse in matching layer to recognition. On the other hand, the face contour features,2D face image features and 3D depth image features are voted to fuse in decision layer. Results show that the recognition performance has been greatly improved by fusion. The more information, the higher recognition performance. |
PDF Full Text Request