High-precision three-dimensional face reconstruction method based on graph neural network

Abstract

The invention relates to the technical field of three-dimensional face reconstruction, in particular to a high-precision three-dimensional face reconstruction method based on a graph neural network. Comprising the steps that a face image is preprocessed into an RGB image with the 64 * 64 pixel size, the RGB image is input into a neural network encoder, and texture parameters, shape parameters and space and illumination parameters are obtained respectively; inputting the texture parameters into a texture decoder to generate a texture map, and inputting the shape parameters into a shape decoder to generate a depth map; pixel coordinates of the texture map are converted into X and Z coordinates in a space and gridding is carried out, the size of a corresponding pixel value in the depth map is used as a Y coordinate, three-dimensional face vertex coordinates are obtained, the attitude is corrected in combination with space parameters, and a face preliminary three-dimensional model is obtained; and through a graph neural network, features between adjacent points are aggregated, and the positions and textures of spatial points are optimized to obtain a smoother and more real face model. The method has a good smoothing effect on the three-dimensional face model, so that the shape, texture and color of the three-dimensional face model are more real.

Description

technical field [0001] The invention relates to the technical field of three-dimensional face reconstruction, in particular to a high-precision three-dimensional face reconstruction method based on a graph neural network. Background technique [0002] The 3D face reconstruction technology extracts information from 2D face pictures and establishes a corresponding 3D face model. At present, most of this technology uses the method based on 3D deformable face model (3DMM) for linear reconstruction, and the linear reconstruction method can be divided into traditional key point parameter fitting method and neural network parameter fitting method; a few methods do not rely on 3DMM model, Direct nonlinear reconstruction using neural networks. [0003] The 3DMM method formula is as follows: [0004] [0005] [0006] where S mean , T mean Represent the statistical average face shape and texture, respectively, s i with e i are the principal components of face shape and exp...

AI Technical Summary

Problems solved by technology

Since the 3D face model contains tens of thousands of vertices, the amount of data that needs to be predicted by the neural network is very large, so this method has a large amount of model parameters, and it is difficult to converge during the training process, and the transfer effect is not good.

Images