Research On 3D Face Reconstruction Technology Based On Generative Adversarial Network

With the widespread application of 3D face reconstruction technology in face recognition,security warning,target tracking and other fields in recent years,research on 3D face reconstruction methods has become a hot topic in the field of computer vision.Although the three-dimensional reconstruction technology has a certain theoretical support and a broad application foundation,how to ensure the authenticity and accuracy in the process of face restoration is still a difficult research.The traditional three-dimensional face reconstruction method needs to fully consider the geometric association of facial features and facial texture structure.The cost of the information obtained in order to complete these associations is extremely high.In addition,the traditional three-dimensional face reconstruction technology requires repeated scanning by external devices such as infrared scanners to obtain target feature information,which requires a lot of economic and time costs,and it is difficult to fuse other modal data information in actual use.The method of reconstructing three-dimensional graphics through two-dimensional images can effectively solve the above-mentioned potential problems in the traditional three-dimensional reconstruction technology.In recent years,more and more researchers have begun to pay attention to and study the method of 2D face image reconstruction 3D face model,especially the emergence and development of the generation of anti-neural network,which provides the 2D face image reconstruction 3D face model.New technology path.In this paper,combining the characteristics of the generative adversarial neural network,with the focus on mining potential features such as cloud points,voxels,and depth in the face image,a three-dimensional face reconstruction method based on the generative adversarial neural network is proposed.The main work of this article is as follows:1.Aiming at the problem of the scarcity of real face data samples,an image acquisition platform was constructed and 2105 portraits of 3105 volunteers were collected to complete the 3D portrait data set,which provided high-quality molding data for subsequent experiments.2.Due to the problems of sparse sample size and limited variability in the publicly available 3D face database(for example,few subjects,small differences in subjects' race and gender),and some problems in the deep network grid convolution The completely solved challenge leads to operator approximation and model instability,often unable to retain the distributed high-frequency components.This paper proposes a three-dimensional portrait reconstruction model(3D-UV-GAN)based on generative adversarial neural networks.The 3D-UV-GAN model is a generative adversarial network model built for 3D face distribution modeling,while retaining the high-low frequency components and shapes of the 3D face.The experimental analysis on a large number of data sets shows that the model can perform well in tasks such as 3D shape expression,generation and transformation.3.Propose a new generation adversarial network 3D-DM-GAN model that can generate high-quality three-dimensional images through clear two-dimensional image expression.At the same time,an end-to-end generative learning framework for constructing 3D portraits and 2D portraits is proposed.The framework first learns 2.5D images(depth maps)that are not easily distinguishable based on real shapes,then completes the 3D reconstruction target from the 2.5D maps,and finally,learns the real textures of the 2.5D maps and generates 3D portrait images.Experimental results show that the frame model can generate more realistic three-dimensional images.