The Study Of 3D Face Reconstruction Algorithm Based On Biocular Stereo Vision

With the development of computer vision,3D face reconstruction technology has great potential in many fields.Due to the low cost and easy operation,3D reconstruction based on binocular stereo vision has attracted extensive attention.After decades of investigation,it has achieved great success,however,the main problem is that it is difficult to reconstruct 3D face model with high accuracy.In order to solve the matching problems of low-texture face images caused by general stereo matching algorithms,this thesis proposed a stereo matching algorithm and a disparity correcting algorithm.Based on the system established in this thesis and the FRGC v2.0 face database,a large number of experiments have been carried out to prove the superiority of the proposed algorithm.The results show that the proposed algorithms improve the matching accuracy of face images,and the root mean square error of the 3D face model reaches millimeter level.The main contents of this thesis are as follows:1.A region-growing stereo matching algorithm based on seed point optimization is proposed.First,the face key points extracted by the cascade regression tree algorithm are used to divide the face into different regions,and the disparity search range of each region is limited according to the one-to-one correspondence characteristics of the face key points in two images,thereby avoiding finding matching points on the global scale.Then,face edge feature points extracted by Canny edge detection operator are matched and used as initial seed points.According to the local shape characteristics of the human face,the method of local surface fitting is used to filter out mismatched seed points and generate a larger number of reliable seed points for region growing.Finally,the stereo matching of the entire face is realized by region growing.The experimental results show that this algorithm can greatly improves the density and accuracy of the seed points,thereby effectively improving the matching accuracy of the face images.2.A dynamic programming disparity correcting algorithm based on bilateral weight is proposed.First,the local surface fitting algorithm is used to detect mismatch points in the facial disparity map.Then the bilateral weight values between each pixel and the effective points in the neighborhood window are calculated in each row of image,the effective points and the correct matching points in the same line of the mismatched points are respectively used to guide dynamic programming in disparity smoothing constraints and global constraints,which realize the disparity correction of the mismatch points in each row of image.At last,the 3D point cloud of the face is calculated by the triangular geometric relationship,and the 3D face model is reconstructed by point cloud meshing and texture mapping technology.The experimental results show that this algorithm can effectively solve the matching problems of low-texture face images caused by general stereo matching algorithms,and reconstruct a more realistic 3D face model.3.A complete 3D face reconstruction system based on binocular stereo vision is implemented.The stereo camera calibration,epipolar rectification,image preprocessing,stereo matching,disparity correction,3D face reconstruction and other steps are completed for constructing the system,the face reconstruction algorithm proposed in this thesis can be used in the system to obtain a better 3D face model.