Research On Stylized Editing Method Of Face Attributes Based On Improved StarGANv2

With the development of artificial intelligence and computer vision,face synthesis technology has been applied largely in a wide range of social fields,which has greatly promoted the development and progress of economic society.As a research hotspot of face synthesis technology,face attribute editing technology has been widely used in artistic life,film and television photography,criminal investigation and other fields.After the creation of GAN(Generative adversarial networks),face attribute editing technology has made great progress and achieved the editing of specific attributes.However,the image generated by this series of methods for attribute editing lacks diversity.Therefore,StarGANv2 separates the attribute style feature and identity feature of face,realizes attribute stylized editing through style transfer,and retains the identity feature of the input face.However,our research found that some images generated by it has poor attribute stylized editing effect and unnatural facial texture details.Therefore,based on the analysis and research of technical theories involved,this paper proposes commensurable solutions to the shortcomings of images generated by StarGANv2.The main work is as follows:(1)A face attribute stylized editing model based on improved StarGANv2 is proposed.First,to improve the feature extraction ability of the generator of StarGANv2,we embed a simplified PSConv module with multi-scale features fusion function in the generator;Secondly,a multi-dimensional dynamic attention module—MDConv is proposed and embedded into the discriminator of StarGANv2 to improve the ability of discriminator to differentiate between true and false images.(2)The effectiveness of the above method is verified through the experiment of qualitative analysis and quantitative comparison.The main contents include:comparative experiment about face images generation of our proposed model and StarGANv2 model on CelebA-HQ dataset is carried out.Results show that our proposed model has better performance on generating images,the attribute editing effect is more beautiful,the facial texture details are more natural,and the values of FID and LPIPS are both improved;In addition,ablation experiment is carried out to further verify the advantage of the method used in this paper.