Scene-adaptive fine three-dimensional face reconstruction method and system, and electronic equipment

Abstract

The invention belongs to the technical field of image processing and mode recognition, particularly relates to a scene-adaptive fine three-dimensional face reconstruction method and system and electronic equipment, and aims to solve the problems of strong model feeling and poor generalization of a reconstruction result of existing three-dimensional face reconstruction. The method comprises the following steps: augmenting face shapes of a training set based on a 3DMM and a graphical imaging model to obtain a plurality of pieces of three-dimensional face data and corresponding images thereof; fitting a three-dimensional variable model for the image corresponding to the three-dimensional face data as an initial shape, and performing virtual multi-view generation based on the image corresponding to the three-dimensional face data and the initial shape to obtain a multi-view image; inputting the multi-view image into a many-to-one funnel network, performing optimization through a loss function with consistent vision to obtain a refined three-dimensional face shape. According to the method, the three aspects of training data construction, model design and scene self-adaption are improved, and the robustness of the model in an unconstrained scene is improved while fine shape reconstruction is achieved.

Description

technical field [0001] The invention belongs to the technical field of image processing and pattern recognition, and in particular relates to a scene-adaptive fine three-dimensional face reconstruction method, system and electronic equipment. Background technique [0002] Most of the current 3D face reconstruction algorithms are based on 3D variable model (3DMM) for shape reconstruction. However, limited by the cost of 3D data collection, most 3DMM models are constructed using only a few hundred scanned point clouds. The span of age, race, etc. is very small, and they are often photographed in a controlled environment. The human face is frontal and has a natural expression. The resulting model has limited expressive ability, and it is difficult to describe all the changes that may exist in the actual human face. [0003] The existing mainstream algorithm is based on convolutional neural network for 3D face reconstruction, and its training generally requires a large number o...

AI Technical Summary

Problems solved by technology

[0005] In order to solve the above problems, that is, in order to solve the shortcomings of the existing three-dimensional face reconstruction method, the reconstruction result has a strong sense of model and poor generalization, the present invention provides a scene-adaptive fine three-dimensional face reconstruction method, system, and electronic equipment

Images