Research Of 3D Face Reconstruction Based On A Single Image

Face shape and surface material is very complicated, many problems in real life in 2d environment cannot obtain a good solve results. For instance in face recognition, face monitoring, etc, the recognition rate of the varieties of facial expression or the matching between positive and lateral is always cannot be improved. The rate and robustness of recognition will be improved if we can artificially synthesize the facial expressions by a 3D model or using 3D information for matching. Currently, the 3D facial model is widely applied in security certificate, film and cartoon, medicine science and many other fields. In recent years, the study of 3d face reconstruction has also made many good results. However get real detailed facial information is very expensive, for example, need to use 3d laser scanner, at the same time, it is difficult to implement due to the environmental factors or the degree of adaptability of the modeling objects. Because the 2D image is very easy to obtain, so there are a lot of researches reconstruct 3D facial models based on 2D images recently. Among them, due to the positive image required conditions rarely so very accessible, so this article will discusses the method of reconstruct the 3D facial model be based on a single photo of the two-dimensional information.This paper first comprehensive summarize and analysis the current shape from shading (SFS) methods, such as the minimization approaches, the local approaches, linear approaches and propagation approaches, we choose the propagation approaches and then proposed a 3D reconstruction arithmetic based on the level set method and energy minimization. We proposed an improved fast marching method to solve the level set problem after the analysis of the traditional level set numerical methods. According to the experiment results it is obviously that the accuracy of solution is enhanced remarkably by using this method. We can get the 3D facial shape using the improved fast marching method which is rather consistent with the original 2D image. But due to the inherent ill-posed of the SFS problem, the final results cannot avoid convex-concave ambiguities. In order to solve this problem, we calculate the flatness of the triangle faces in order to search the boundary of convex-concave ambiguities, so we can determine the concave areas. The error of this method is small by the initial experiment. Then we reconstruct the concave areas by energy minimization based on the centerline and the boundary of the surface, the experimental results show that the 3d facial model which we get eventually is similar as the original 2d image, the authenticity is quite high.