Research On Face Attribute Editing Based On Generative Adversarial Network

Face attribute editing refers to changing certain attributes in a facial image while preserving other attributes such as hairstyle,skin color,and age.This technology has been applied in various fields such as beauty and hairdressing,intelligent movies,and facial dataset expansion.It not only enhances existing facial beauty techniques but also plays a crucial role in maintaining the accuracy of the target attribute image and the invariance of irrelevant attributes.However,despite the significant contribution of Generative Adversarial Networks to the development of face attribute editing,existing algorithms still have some shortcomings.Some algorithms focus on the completeness of image information and cannot ensure the independence between different attributes,leading to problems such as deformation of nonfacial areas and changes in background information during attribute editing.Furthermore,most algorithms currently rely on reference images for face attribute editing,and there is limited research on face attribute editing algorithms that combine image and text,with issues such as missing image details and unsatisfactory editing results.(1)A new attribute difference method is proposed in this article,which only considers the attributes to be edited.The difference between the target attribute vector and the source attribute vector is used as the input to the encoder-decoder to enhance the flexible editing capability of the attributes.The attribute difference vector can focus more on the semantic information of the location area where the attribute to be edited is located,to ensure that the irrelevant areas of the attribute are not changed.Meanwhile,the reconstruction loss is added to the constraint to ensure that the image reconstruction results are consistent with the input image space distribution,thereby constraining the quality of image reconstruction.The attribute classification loss is also added to guide the generator to establish the association between attribute labels and generated images,ensuring the performance of correctly editing the target attributes.Experimental results show that the proposed attribute difference method can accurately edit various facial attributes while effectively preserving the irrelevant attributes.(2)A text-based control method for facial hairstyle editing is proposed in this article.A new hairstyle editing interactive mode is proposed that can edit hair attributes based on userprovided text or reference images.First,the text encoder and image encoder of CLIP are used to encode the input conditions into 512-dimensional conditional embedding vectors,which are used as inputs to the mapping network.To achieve the decoupling of hairstyle and hair color,three identical sub-network mapping networks are proposed to correspond to different semantics in the latent space,in order to improve the representation independence of attribute sub-vectors in the latent space.To ensure that other irrelevant attributes remain unchanged in the editing task,three types of losses are introduced: 1)text manipulation loss to ensure the similarity between the edited result and the given text description;2)image manipulation loss to guide the hairstyle or hair color to be transferred from the reference image to the target image;3)attribute preservation loss to keep the irrelevant attributes(such as identity and background)unchanged before and after editing.Experimental results show that this method can better preserve the details of hairstyle and hair color,as well as the realism of the image,and has significant effects in facial attribute editing tasks.