Face Attribute Editing Based On Mask And Attention Mechanism

Face attribute editing is a method to change some attributes of face image,such as makeup,hairstyle,facial expression,while keeping other attributes unchanged.The goal is to improve the accuracy of attribute editing and the quality of generated image.In recent years,the development of generative network,especially the generative confrontation network Gan,has gradually matured,which promotes the development of face attribute editing.However,there are still some problems in the existing algorithms:first,it is difficult for the existing algorithms to achieve effective dissociation when editing a single attribute,and there is a phenomenon of entanglement between attributes.Editing an attribute often leads to significant changes in irrelevant attributes;Second,most algorithms seek to edit the face attributes,but ignore the pursuit of attribute styles,such as smiling mouth.Thirdly,the image generated by face attribute editing often has the problems of background change,artifact and low quality;Fourth,in order to improve the editing effect,the existing algorithms are highly abstract to the image features,which makes the image identity information lost,resulting in the generated image is not natural and real enough to achieve the ideal editing effect.To solve the above problems,this paper makes a series of improvements on the basis of the existing algorithm(1)To solve the problem of entanglement among attributes,this paper proposes a face attribute editing algorithm based on mask guidance.In this algorithm,the attribute mask is introduced to filter the high-dimensional spatial features of the image,so as to accurately locate the attribute region to be edited and avoid the change of irrelevant attributes.In addition,the attribute label and reference image are fused in the model to guide the editing,and the unpaired face image is used to provide users with a variety of attribute templates.By combining the advantages of the three conditions,the algorithm can achieve diversification and integration More accurate face attribute editing effect.(2)Aiming at the problem of low quality of generated image and easy to cause irrelevant changes caused by attribute editing,this paper introduces spatial attention mechanism to assist network to generate attribute mask corresponding to image.The attribute mask obtained by attention network learning contains richer semantics and can generate more real and natural editing results.At the same time,aiming at the problems that may occur in the previous stage,such as the distortion of the generated image background and the degradation of the generated image quality,the model uses the attribute mask to fuse the edited region of the generated image with the irrelevant region of the original image to get the final generated image.This operation improves the training speed of the model and ensures the invariance of the image background and the irrelevant region,It can generate high quality editing image.