Face Attribute Editing Method Based On Generating Adversarial Networks

Face attributes refer to some visually understandable semantic features of a face image,such as glasses,gender,age,etc.Face attribute editing refers to tasks such as adding glasses,changing gender and age,which operate on a single or multiple attributes specified in a given face image.The purpose of face attribute editing is to accurately change specified attributes while ensuring that regions unrelated to the specified attributes do not change.With the continuous development and maturity of deep learning,especially the generative adversarial networks,great achievements have been achieved in face attribute editing,but the following problems still exist in the field of face attribute editing:(1)Existing methods are prone to losing information during the up and down sampling process,and the edited images are prone to unnatural phenomena such as artifacts and distortions,making it difficult to achieve realistic and natural editing effects;(2)existing methods are difficult to achieve good attribute de entanglement effects,and when editing target attributes,it is still easy to cause unrelated attributes to be tampered with,making it difficult to achieve accurate and controllable editing effects.In order to solve the above problems,this article has made the following improvements based on existing methods:(1)Aiming at the problem of information loss during up and down sampling,this paper proposes a facial attribute editing method ASCGAN based on attention skipping connection.Based on the architecture of the encoder and decoder,in order to capture long-distance dependencies during feature extraction,this paper introduces a self attention module in the shallow layer of the encoder and decoder,thereby improving the ability to edit global attributes;on the basis of U-Net skipping connection,this article introduces an attention skipping connection module between the deep layers of the encoder and decoder,using skipping connections to enrich the detailed information used in up-sample on the decoder,and filtering out the information related to the target attribute in the detailed information through an attention mechanism,retaining only the information unrelated to the target attribute,thereby improving the quality of image generation without negatively affecting editing capabilities.(2)To solve the problem of attribute entanglement,this paper proposes a face attribute editing method FSD-ASCGAN based on semantic direction and deviation correction based on ASCGAN.In order to avoid attribute entanglement caused by the overlap of global attributes with large coverage and unrelated attributes in the region during style migration,this paper introduces a semantic transformation module based on semantic direction,which can explain the potential representation embedded in the potential space along the target semantic direction to complete the task of attribute editing,enabling the model to better suppress the change of unrelated attributes when editing global attributes;in order to avoid the possibility of over fitting or under fitting in the training of attribute classifiers in the presence of data bias,and not being able to impose accurate constraints on unrelated attributes,this paper uses the strength differences of potential representations in the semantic direction of each unrelated attribute before and after editing as constraints,and uses the correlation between attributes to impose varying degrees of constraints on unrelated attributes,making the model aware of the uneven distribution of data,thereby achieving better performance in disentanglement.