| Face 3D modeling is a computer technology that has been developed for a long time,and has been widely used in medical,gaming,movie entertainment and other fields.The traditional method for obtaining a 3D face has high cost and complicated process,which limits the application scenarios of the 3D face to a certain extent.With the further development of computer vision,the technology of 3D face reconstruction from 2D images is gradually taking shape.This method has simple preconditions and is easy to achieve full automation.However,the 3D model reconstructed by this method still has problems such as inaccurate geometric shapes and texture mapping effects that are easily affected by lighting,and there is still much room for improvement.This paper studies the main problems and key technologies in the process of 3D face reconstruction,and proposes a set of algorithms that can use deep learning to reconstruct 3D faces from a single image stably.Firstly,in view of the fact that highlights often exist in real face images,an improved face highlight removal algorithm is proposed.We designed a "segment-process-fuse" highlight removal pipeline,and upgrade the algorithm based on the dichromatic reflection model and bilateral filter.We propose the idea of “diffuse compensation” and a method of estimating the compensation to solve the problem of “dark spots” in the traditional highlight removal algorithm.Then,on the basis of the existing 3D face reconstruction method based on UV position map,we improved the encoder-decoder network model with the idea of grouped convolutions.And at the same time,on the premise of ensuring high accuracy,we have greatly reduced the amount of network parameters and calculations.After detailed comparison experiments and analysis of the results,the 3D face reconstruction algorithm proposed in this paper has low computational complexity and high accuracy,and can stably reconstruct a 3D face model with high-quality texture from a 2D face image with highlights. |
PDF Full Text Request