Research On Method For Face Reconstruction Based On Deep Learning And System Implementation

Three-dimensional face reconstruction is a challenging problem in computer vision and is of great value for applications in face recognition,animation film and video and game production.With the development of deep learning techniques,the neural network-based 3D face reconstruction techniques are able to learn complex features of faces and reconstruct 3D face models.In this paper,an end-to-end 3D face reconstruction system is designed based on the face representation method of UV position map.In order to enhance the robustness of the face reconstruction model,this paper investigates from a feature perspective and proposes a 3D face reconstruction model based on a deep crossover network.On the premise that the 3D face is represented in UV position coordinate space,the features are processed by coding and decoding.By downsampling the features,the computational effort can be reduced and a feature map with a large perceptual field can be obtained at the same time.This paper introduces a deep crossover network at the end of the downsampling,which,by introducing weights and bias components,can process the original features and form new feature vectors,making the newly generated features correlated with the original features.The method was validated on a dataset and proved to be able to obtain good results.A dense grid-based 3D face reconstruction algorithm is proposed to address the problems of sparse areas in the face reconstruction process and the lack of rich representation at the bend of the face after reconstruction.The triangular mesh algorithm is able to provide a good surface fit to the face region in the face reconstruction process,and the surface position contains more triangular facets.In this paper,we use Lagrangian interpolation to resample the points in the mesh and reconstruct the triangular mesh to obtain a more dense face model.Experiments are conducted to compare the reconstruction effect before and after mesh densification to prove the effectiveness of the proposed method.Finally,a 3D face reconstruction system is designed,which can reconstruct 3D faces from 2D face images.