Research On The Editing Method Of Face Arbitrary Attributes Based On Generative Adversarial Nets

Face arbitrary attribute editing is a hot topic in computer vision and generation model,and it is an application technology based on face recognition and generation.The main purpose of face arbitrary attribute editing is to control the arbitrary attribute representation of the face image according to the given face attribute tag value,so as to obtain the fake face image that meets the specified attribute requirements,and at the same time,the image areas in the generated fake face image that do not need to be changed should be consistent with the original image.In recent years,face arbitrary attribute editing has been widely used in medical beauty,criminal investigation,entertainment,facial recognition and other fields,and it has attracted more and more attention.At present,the face arbitrary attribute editing model is usually implemented in combination with the encoder-decoder and the generative adversarial nets,but there are problems that the generated image is not real enough,the attribute editing of the face image is not accurate enough,and the fine-grained control ability of the generated image face attribute is weak.This thesis focuses on the problem of editing arbitrary attributes of faces,and the main work includes:(1)Aiming at the information redundancy problem caused by the use of jump connections between encoder-decoder to transfer image features in the face arbitrary attribute editing model and the problem that the model is difficult to control the degree of attribute change.The thesis designs a face arbitrary attribute editing network ISTSA-GAN which combines independent selective transfer unit and self-attention mechanism.The network combines the independent selective transfer unit and the encoder-decoder as the generator.The encoder can obtain the potential features of the image,the independent selective transfer unit selectively converts the potential features of the image obtained by the encoder,and the self-attention mechanism is introduced into the decoder to establish the long-distance dependence across image regions,and then the attribute interpolation loss and source domain confrontation loss are added to constrain the training of the model.Experimental results show that this method can improve the ability of attribute editing and detail saving,and can enhance the ability of fine-grained control of editing attributes.It outperforms Star GAN,Att GAN and STGAN in attribute editing accuracy and quality of generated images.(2)Aiming at the problems that the input of the decoder in the face arbitrary attribute editing model is redundant and the filtering of information by the single-layer gating in the selection conversion unit is not accurate enough.The thesis designs a face arbitrary attribute editing network based on improved U-Net structure and double-layer gated selective transfer unit.In this network,double-layer gating is used to adaptively learn the retention of existing memory information and hidden information in the past.At the same time,part of the output of the first-layer double-layer gated selective transfer unit is used as the input of the decoder,and then the attribute interpolation loss and source-domain confrontation loss constraint model training are additionally added.Experimental results show that this method can improve the ability of editing attributes,and further enhance the ability of fine-grained control of editing attributes.