An end-to-end 3D face reconstruction method based on a single image

Abstract

An end-to-end three-dimensional human face reconstruction method based on a single image, mainly comprising the following steps, step 1: prepare a reference three-dimensional human face model as a reference template, and the reference three-dimensional human face model is in an expressionless state; step two: A triangular grid parameterization method is used for the reference template, and the reference template is represented in two dimensions to obtain the UV coordinate value corresponding to the reference template, and the UV coordinate of the reference template is used as a reference position to represent a space. The effect achieved can use the position map to directly represent the 3D face model, and use the neural network to directly predict without going through the 3D deformation model, thereby realizing the end-to-end reconstruction process, and the accuracy of the reconstruction model is not limited.

Description

technical field [0001] The invention relates to the field of face recognition, in particular to an end-to-end three-dimensional face reconstruction method based on a single image. Background technique [0002] 3D face reconstruction based on a single image, that is, to reconstruct a 3D model of the face from the information of a single image. Since face recognition is an important task of artificial intelligence in the field of intelligent perception, it also has great practical value, and 3D face recognition has the characteristics of invariant posture and illumination, making 3D face reconstruction an important task in computer vision and computer graphics. A research hotspot in the field of science. At the same time, although structured light and binocular cameras can restore the 3D information of the face very well, in most application scenarios, only a single image of the target face can be obtained or a monocular camera is used for data collection. Recovering accurat...

AI Technical Summary

Problems solved by technology

But there are still shortcomings: the reconstructed 3D face model is limited by the 3D deformation space, and the accuracy is not enough

One representation method is to vectorize the vertices of the 3D model. This method loses the relationship between the points of the 3D model, making it difficult for the model to converge, and the accuracy of the reconstruction result is not high

Another method is to use voxels to represent the 3D model, that is, use the 0, 1 of the 3D array to indicate whether it is part of the face model. This representation method retains the spatial relationship between the 3D model points and points, but the representation The large size of the method makes the neural network model more complicated and takes a long time to run, and also limits the accuracy of the reconstructed 3D model

Images