Research On Face Attribute Editing Algorithm Based On Attention Mechanis

With the rapid development of computer vision,facial attribute editing has been widely used and expanded in practical scenes such as image beautification and video beautification.Facial attribute editing refers to the operation of single or multiple attributes of face image to generate a new face that meets the target attributes while retaining other details.Generative Adversarial Networks promote the development of facial attribute editing,but there are still deficiencies.In order to solve the problem of multi-attribute entanglement in feature space in the facial attribute editing task of Generative Adversarial Networks,this paper conducts in-depth research on the facial attribute editing algorithm based on the attention mechanism.The main research results are as follows:1.A Cross Channel Self-Attention based Facial Attribute Editing Method(CCA-GAN)was proposed.The network weighted the importance of multi-channel features and completed pixel-level feature alignment and transformation to reduce the impact of irrelevant attributes when editing target attributes.The evaluation results show that CCA-GAN has higher quality face image generation results on Celeb A dataset.2.A Facial Attribute Editing algorithm based on Dual Cross Channel Self-Attention(DCCA-GAN)is proposed.Aiming at the difficulty of decoupling multiple face attributes in feature space,dual cross channel self-attention is introduced to filter and combine multidimensional vectors containing a large number of facial details,so as to decouple multi-attribute features in feature space.Effectively improve the representational independence of attributes.By combining absolute attribute coding with Selective Transfer Units(STU),more abundant characterization information can be obtained.Compared with the existing methods,DCCA-GAN can effectively reduce the problems of background change and face irrelevant deformation on Celeb A dataset,and generate target images with higher quality and more accurate attribute editing style.At the same time,it also achieves good results in style transfer task.