A Facial Expression Synthesis Method Based on Generative Adversarial Network
Bin Xu, Weiran Li, Qing Zhu · 2022
Recently, machine learning, especially the emergence of generative adversarial networks (GANs), has further enhanced the robustness and realism of facial expression conversion models. However, most models have flaws such as fuzziness in the details. Based on this, this article mainly studies the facial expression synthesis method based on GANs. Firstly, we created a dataset containing 127,616 expression annotations suitable for the study of facial expressions. The dataset has been tested on mainstream models with good generation results. Secondly, we propose a GAN network structure named SRFEGAN with a super-resolution synthesis module. This module helps solve the artifact problem in the process of image conversion. Experimental results on our dataset show that the average recognition accuracy rate of the generated images is 63.76% and the Frechet Inception distance (FID) is 36.581. This shows that our network can accurately synthesize the facial expression image of the subject, and the image quality is better.