3D Face Representation And Reconstruction Based On Multiscale Graph Convolution Neural Network

The effective representation and reconstruction of 3D faces has a wide range of applications in computer vision and graphics.The representation of the 3D face is to obtain the high-dimensional feature data of the 3D face through a certain algorithm.The reconstruction is to reconstruct the corresponding 3D face from the high-dimensional feature data through the algorithm model.Most existing linear representation algorithms are not effective in reconstructing high quality 3D face data,especially for facial details,and the latest nonlinear representations are not suitable for actual 3D shapes.The representation and reconstruction of 3D face can provide technical support for face recognition,machine emotion expression and other directions,and provide reference for other directions such as 3D human body drive and scene 3D reconstruction.In order to solve the effective representation and reconstruction of 3D face,this paper proposes a multi-scale autoencoder neural network model based on graph convolution to deal with the potential representation and reconstruction of 3D face.This paper proposes that the autoencoder neural network model uses graph convolution,which is a neural network module suitable for graph structure data.It solves the problem that the traditional convolutional neural network is difficult to extract features effectively.Combined with the sampling algorithm of 3D mesh,the feature extraction of 3D mesh can be performed on multiple scales,which is similar to the traditional 2D data pooling operation.The model in this paper is also capable of variational training.The trained network can be used to generate more high-quality 3D face data and data enhancement for other tasks.The innovations and achievements of this paper are as follows:1.A multi-scale automatic encoder model based on graph convolution is proposed.The feature convolution is used to extract the features of 3D face data effectively.Combined with the sampling algorithm of mesh data,multi-scale feature extraction is realized.The traditional convolutional network model can not be effectively generalized to the problem of 3D space,and the potential layer representation and reconstruction of the face model are realized efficiently by using the automatic encoder structure.The model in this paper is also applicable to most other types of 3D models,not limited to face models.2.The 3D mesh data does not need to perform complex coding process.The network model proposed in this paper can directly input the 3D mesh raw data,without the need of complex data representation coding process,that is End-to-end input and output are achieved.3.The variational generation model of 3D face mesh is realized.The training of variational generation model needs to weigh the weight between relative entropy and reconstruction error in loss function.This paper adopts self-growth weight setting algorithm to realize The stable training of the network model can generate high quality 3D face models.