Three-dimensional face reconstruction method

Abstract

The invention discloses a three-dimensional face reconstruction method which comprises the following steps of S1, positioning two-dimensional key points of a face image to be reconstructed; S2, positioning a three-dimensional key point coordinate in a three-dimensional model corresponding to the two-dimensional key points in the S1; S3, weighting-calculating shape reconstruction coefficients of a three-dimensional face and imaging model parameters by utilizing subunits of the key points in the S2; S4, synthesizing the shape of the three-dimensional face by utilizing subunits of a three-dimensional deformable model; and S5, extracting texture of the three-dimensional face by utilizing the imaging model and face symmetry. By means of the three-dimensional face reconstruction method, good reconstruction results of face images in any postures can be obtained, the computing is fast, and nonlinear optimization can converge quickly. Compared with the dense reconstruction algorithm in the prior art, the speed is improved by more than one order of magnitudes, and the performance is stable.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to a three-dimensional face reconstruction method. Background technique [0002] 3D face reconstruction is an important research topic in computer vision and computer graphics, and is widely used in the fields of face recognition, human-computer interaction, and 3D animation. The classic method of 3D face reconstruction is the 3D Morphable Model (3D Morphable Model), which uses the linear combination of basis vectors to reconstruct the 3D shape of the face x model and texture t model ,Right now [0003] x mode l = x ‾ + Sdiag ( σ ) c ...

AI Technical Summary

Problems solved by technology

First of all, the small number of key points can easily lead to overfitting. Secondly, due to the limited number of samples for training the 3D deformation model, the small sample problem will lead to limited generalization of the synthetic model. The 3D deformation model is a linear model, and its expressive ability Limited by the training library

Because the cost of 3D acquisition and processing is relatively high, the existing database currently has no more than three or four hundred people at most, which is far from the number of tens of thousands of face vertices. Therefore, the problem of small samples may lead to a larger generalization error of the model

In addition, if the camera projection matrix used for reconstruction is inconsistent with the real projection matrix that produces 2D coordinates, additional systematic errors will be caused

Images