3D Face Recognition Based On Contour Lines And SIFT Descriptors In Circular Neighborhood

Face recognition has been a promising approach in the area of biomatric authentication for its natural, user friendly and non-disturbing. Although 2D face recognition has achieved reliable recognition performance over the past decades, shortcomings of susceptible to variations in illumination, posture and facial expressions still hinder its improvement. Adequately represent the human face with richer information and overcoming the insurmountable problem of 2D face, 3D face recognition has become a favourite. This thesis discusses itself to the work of 3D face recognition based on contour line and Scale Invariant Feature Transform(SIFT) feature descriptor in circular neighborhood. The main work is as follows:(1)In order to get positive point cloud of 3D face, it respectively providess the procedure of nose location, face segmentation, rough registration for pose correction with PCA and accurate registration with ICP algorithm. After then convert the registed point set to 2.5D depth map and 2D intensity map. All the preprocessings procesures are to lay well foundation for the feature extraction steps next.(2)Aimed at shortcomings of the high feature vector dimension and long time consuming during the feature matching with traditional SIFT descriptor, an improved algorithm is proposed. It uses a circular neighborhood in which the feature point as a center insteads of the rectangular neighborhood in the feature description. At features point matching stage, according to the prior knowledge of face images to eliminate some false matching points. Experimental results show that, the proposed algorithm can effectively reduce the feature matching time, and get a recognition rate of 90.5%.(3)To improve the efficiency of 3D face recognition system, a cascade classifiers based on contour lines and SIFT features in circular neighborhood is designed. Meantime adopt the score-level fusion strategy on depth map and intensity map to improve the recognition performance. The experimental demonstrates that the proposed recognition system can reach an accuracy of 92.7%.