Research On Facial Expression Synthesis

Facial expression synthesis is one of the important research directions in the field of computer vision.In recent years,research on facial expression synthesis has made significant progress,and facial expression synthesis technology has gradually been applied to various fields of production and life.In order to generate high-quality facial expression images,scholars have conducted extensive research in different directions and proposed a series of facial expression synthesis models.These generation models can be roughly divided into three categories: Feature-Guided models,Image-to-Image models,and Encoder-Decoder models.After in-depth research and analysis of these models,it is found that there are still many challenges in the current facial expression synthesis models,such as(1)the generation of facial images with multiple expressions is still the focus of research,(2)the generated facial expression images should present the target expression as accurately as possible,and(3)the parts of the generated facial expression images that are not related to the expression should be consistent with the original facial image.This article studies the above three issues,and the work is as follows:(1)A detailed analysis of facial expression attributes and facial expression image synthesis models has been conducted,and an end-to-end generative adversarial network that can meet the mutual synthesis of multiple expressions is designed.The network has only one generator and one discriminator,which effectively reduces the complexity and training time of the neural network.(2)Using the feature disentanglement algorithm,by disentangling the extracted face features,the mutually orthogonal expression-related code and expression-irrelated code are decomposed,and then the target expression code is introduced for editing.And the method of controlling the introduction of information is adopted,and the expression transfer is performed on the input facial expression image by introducing the edited code at the feature level.On the Ra FD dataset and the Celeb A-HQ dataset,it is proved that the real facial expression image synthesized by the real facial expression image synthesis model designed in this paper can present accurate target expressions,while keeping the irrelevant part of the real facial expression intact.(3)A targeted improvement has been made to the network model to enable it to generate animation facial expression images based on the expressions of reference images in animation facial expression dataset.Through the animation style transfer model,the animation facial expression image Ra FD-car dataset and Celeb A-HQ-car dataset were produced,and it was verified that the animation facial expression image synthesized by the animation facial expression image generation model designed in thesis can present accurate reference expressions,while keeping the irrelevant part of the real facial expression intact.