Research Of 3D Face Reconstruction Based On Binocular Vision

At present,3D face model is widely used in the field of face recognition, expression recognition, video production, etc.3D face modeling has become a hotspot in machine vision. Compared with 3D face modeling methods using laser or structure light scanning technique,3D face modeling method based on binocular vision has gained wide attention for its superiorities, such as the system has advantages of easy configuration and low price, data acquisition process is harmless to the eyes, the movement of the face doesn’t influence the reconstruction accuracy of the model, etc. In this paper, in-depth study is carried out on stereo matching and face surface reconstruction, the core link in binocular stereo reconstruction process.The characteristics and innovations of this paper are listed as follows:(1) In order to further improve the precision of stereo matching, a novel dense matching algorithm of face stereo image pair is presented by deep analyzed the causes of matching ambiguity. Firstly, high-precision original matching points are achieved by improved ALSC matching algorithm with Gaussian weighting. Then these points are used as the seeds, WSSD is used as the similarity measure operator, the region growing algorithm is conducted. Some new match optimizing strategies are added to eliminate matching ambiguity during the region growing process, high-precision dense matching points are achieved. The experimental results show that the proposed method can achieve more accurate and dense matching points than other similar methods.(2) Aiming at the problem that face surface, achieved by present 3D face surface reconstruction methods based on parametric surface, is so smooth that it losses the face detail features, the paper proposes a 3D face surface reconstruction method based on subdivision surface. Firstly, the face point cloud is triangulated by the Delaunay triangulation technique, and thus initial face surface mesh is achieved. By this time, because the number of triangles in this mesh are limited, the initial face surface always seems rough and angular. So, in order to get a smooth and delicate 3D face model, a smooth subdivision surface is generated by using Loop algorithm to subdivide the initial face triangular mesh. Finally, a realistic 3D face model is reconstructed based on this subdivision surface. Experimental results show that the proposed method can fully use the properties of subdivision surface to achieve a smooth and delicate 3D face model. In the meantime, the achieved model also contains lots of face detail features.Finally, a binocular stereo reconstruction system has been designed and established to verify the validity of the proposed modeling method. The theoretical analysis and experimental results both show that the proposed method is effective and robust.