Research On Expression Generation Based On Deep Learning And 3D Reconstruction

Since the development of computer technology,human-computer interaction has become one of the main research contents in the field of computer.With the rapid development of artificial intelligence technology,more and more achievements of human-computer interaction have been used in life,and simultaneously,it drives a lot of research and application related to human face in machine vision.Generative Adversarial Networks(GANs)have been widely used in text,images and other fields since they were proposed in 2014.more and more scholars use its excellent generation ability to conduct research on human faces and have achieved some results.However,in recent years,it has been found that generative adversarial networks are prone to underfitting in the case of limited samples and fail to achieve good generation effect.At present,the number of face dataset samples from Asia,especially Chinese,is rarely in public datasets,which greatly affects the effect of Asian facial expression generation and reconstruction.Aiming at the problem of training on limited data,scholars have proposed a transfer learning method,which is about applying the "knowledge" learned on tasks with sufficient data to different but relevant problems.To a certain extent,it alleviates the problem of neural network performance degradation due to the lack of data,but in the training process,problems such as overfitting are still unavoided because of the limited data.Therefore,the simple use of transfer learning method can't completely solve the problem.In order to solve this problem,this paper proposes a method combining transfer learning and feature map regularization on the basis of generation adversarial network,which can generate and reconstruct expressions from limited data and avoid overfitting.This paper discusses the existing methods of facial expression generation and reconstruction,analyzes the selected generation adversarial network in detail,and states the principle and process of image generation and discrimination.Then,in view of the limited 3D facial expression and difficulty in collecting,a method of generating2 D expressions and then reconstructing them is adopted to generate the final 3D facial expressions.First,transferring pre-trained generative adversarial network with Celeb A face dataset to target network by the way of freezing the weight of low-level convolutional layer and fine-tuning the weight of high-level convolutional layer.At the same time,use the limited Asian expression dataset to generate Facial expressions after transferring.Feature map regularization and auxiliary classifiers are introduced to improve the generation effect,reduce overfitting and generate expressions.Finally,the pretreatment images are input to 3D reconstruction neural network for generating and establishing 3D expressions dataset of Asian faces.For displaying the generation of facial expressions,the demonstration system of generating is designed on the Qt platform.