Research On 3D Face Reconstruction Method Based On Binocular Stereo Vision

The development of image acquisition equipment and the advancement of software processing technology have promoted the development of the field of stereo vision,and face reconstruction as a research hotspot in stereo vision has attracted much attention.Compared with the two-dimensional human face,the reconstructed three-dimensional human face is more robust to the facial posture and illumination changes.It has been widely used in face recognition and virtual reality.Three-dimensional face reconstruction based on binocular stereo vision involves the study of theoretical methods such as face image acquisition,stereo matching,disparity map calculation,depth map estimation,point cloud processing and other links.The data obtained in each link has a direct impact on the effect of 3D face reconstruction.In this thesis,the binocular face image is taken as the research object to study the problem of reconstruction process from binocular image to 3D face point cloud in 3D face reconstruction.Aiming at the problem that the face area point set is susceptible to background interference,based on the positioning of the face key points,the key points of the outer contour of the face are used to establish a convex hull to determine the compact face area to obtain the effective area of the face.A new method for establishing face area correspondence is proposed.Using the left and right face areas as references,feature extraction is performed using the SURF operator,and a correspondence is found in another face area.By fusing key points of the face,the number of corresponding points in the face area is increased.Use the corresponding results to calculate the disparity map.The triangulation method is used for depth estimation to obtain 3D point cloud information.For the reconstruction of the initial three-dimensional point cloud,there are a lot of redundancy problems.In the process of binocular image reconstruction,a method of using the BFM mean face model to constrain the face point cloud is proposed,which is different from the reconstruction method of the deformation model by searching for the correspondence between the two point clouds,the face point cloud model is transformed to the BFM face model,and the shape of the face point cloud is constrained by the shape information of the BFM model,thereby filtering the face point cloud model Denoising.Finally,the correspondence between the face point cloud and the BFM model and the particle swarm optimization method are combined to smooth the surface of the face point cloud to obtain a smoothed point cloud model.