Research On Facial Expression Generation Based On Generative Adversarial Networks

Facial expression generation involves generating facial images with expressions using expression calculation methods.It has wide applications in human-computer interaction,film production,and data augmentation.With the rapid development of deep generative models in image generation,facial expression generation based on deep generative models has received extensive attention from researchers.Among many deep generative models,Generative Adversarial Networks(GANs)have stood out with excellent generation results.Existing GANbased facial expression generation methods aim to solve two problems: the overlapping and blurred phenomena in generated facial expression images,and the issue of generating facial expressions from multiple angles.In this study,we propose two GAN-based methods to achieve high-quality facial expression generation from two perspectives: improving facial expression image quality and generating facial expressions from multiple angles.The main contributions of this study are as follows:1)To address the problems of overlapping,blurring,and lack of realism in generated facial expression images,we propose a GAN model named Group Residuals with Attention mechanism-GAN(GRA-GAN)to generate high-quality facial expression images.First,we embed a hybrid attention mechanism into the generator network before downsampling and after upsampling to adaptively learn key region features and enhance the learning of critical regions in the image.Second,we incorporate the idea of grouping into residual networks and design a group residual block with a hybrid attention mechanism to achieve better generation results.2)To address the low-quality problem of facial expression generation from multiple angles and head poses,we propose a Dense Group Convolution-GAN(DGC-GAN)for generating facial expressions from multiple angles.First,we design a dense group convolution generator with facial key feature point location to alleviate the impact of head poses on facial expression generation.Inspired by the dense convolution neural network,each layer of the generator is connected to all previous layers to improve the network’s feature learning ability.Second,we incorporate group convolution into residual blocks to further improve the efficiency and performance of the model.Finally,we design a multi-branch structure discriminator to enhance the global view and increase the constraint on average pixel values to ensure the correctness of the generated facial expression image angle and content and improve the image quality.Both models are evaluated on the Ra FD public dataset.The experimental results show that both GRA-GAN and DGC-GAN models have excellent performance in improving facial expression image quality and generating facial expressions from multiple angles,surpassing related methods in qualitative evaluation and quantitative analysis indicators.