Research On Facial Expression Synthesis Based On Generative Adversarial Networks

Facial expression synthesis is widely used in face recognition,animation and human-computer interaction.However,expression synthesis is a challenging task due to the non-linear changes in facial expressions.The superiority of Generative Adversarial Network(GAN)in image synthesis can solve the problems of traditional methods to some extent.The existing methods are mainly for expression category synthesis,but lack of fine-grained control of expression intensity.This paper studies mostly the different intensity of facial expressions and synthesizes realistic facial images with different intensity of facial expressions based on GAN.This paper mainly proposes two methods:1)expression synthesis algorithm integrating Attention mechanismWe manually divide the dataset into different intensity categories according to the varying intensity of the expression.Based on the Conditional Generative Adversarial Network,the algorithm achieves the synthesis of different strengths in a supervised way.We use both channel and spatial attention mechanisms in the network,which can focus on facial areas where expression intensity is related.To make the generated image as close as possible to the real target,we use the pixel loss to minimize the difference between the ground-truth and the generated image.Besides,the discriminator can distinguish true and false at the same time can also detect the intensity of expression.We conduct a series of experiments on three datasets,and the results prove the effectiveness of our method in synthesizing facial expression images of different intensity.2)expression synthesis algorithm based on multi-scale feature fusionFeatures between different scales contain different levels of information.The lower features focus on local details,while the higher features are general global information.We weight the features of different scales in an attention-based fusion method to obtain more comprehensive feature information.This method contains both low-level detail information and high-level global information,reducing the problem of detail loss in the process of image generation.At the same time,the attention fusion method makes the network pay more attention to the parts related to expression changes when multi-scale features are fused.A series of experiments verify the effectiveness of the algorithm.