Facial Attribute Editing Based On Multi-attention Mechanism And Generative Adversarial Network

Facial attribute editing has always been a research hotspot in the field of deep learning and computer vision,and there are urgent research needs in the fields of human-computer interaction,film and television entertainment,and criminal investigation.At present,most of the mainstream facial attribute editing networks are based on sparse coder-decoder structures and full convolutional networks,and it is difficult to achieve a balance between the attribute editing and detail retention capabilities.In addition,the coupling relationship between attributes is also a difficult problem that restricts attribute editing tasks.In response to the above problems,this paper proposes a Multi-attention U-Net Generative Adversarial Network(MU-GAN).Firstly,the problem of irreversible loss of detailed information caused by the asymmetric sparse structure of the generator is addressed.The asymmetric convolutional coder-decoder in the traditional generator is replaced with a U-Net-like coder-decoder to reconstruct the coder-decoder information stream.This ensures that the information at both ends of the coder-decoder are symmetrical and avoid the loss of information due to the sudden decrease in the number of channels of the decoder.The features of the encoder at each level are used to supplement the features of the decoder branch to enrich the image details.MU-GAN can retain more image details and generate more detailed and more realistic reconstructed images.Secondly,it is difficult to achieve a balance between model attribute manipulation ability and detail retention ability.In the coder-decoder,an Attention U-Net Connection(AUC)based on the additive attention mechanism is introduced to selectively transfer the detailed features of the encoder and make them complementary to the features of the decoder.AUC helps the generator distinguish the attribute editing area and the non-attribute editing area,and selectively supplements the detail information of the non-attribute editing area,which not only enhances the ability of image detail retention,but also promotes the attribute editing ability of the model.Thirdly,for the problem of limited receptive field of convolution kernel of full convolutional network and poor network feature extractio n ability.A self-attention convolutional complementary layer is constructed.The self-attention mechanism is used to enhance the model’s ability to model long-range and multi-level dependencies,strengthen the geometric constraints of the generated image,improve the model’s feature extraction ability and attribute decoupling ability,and prevent the linkage between attributes caused by the coupling relationship.Finally,We study the method of Multi-attention U-Net Generative Adversarial Network.Through experiments on the Celeb A dataset,it is found that MU-GAN can simultaneously promote detail retention and attribute editing capabilities,and can strengthen the model’s attribute decoupling capabilities.MU-GAN surpasses the current mainstream model in terms of attribute classification accuracy and detail retention,with an average accuracy rate of 89.40%;it obtains a result of 32.53/0.962 in the image reconstruction quality index PSNR/SSIM;in addition,the qualitative experimental results show,MU-GAN’s generated images are more delicate,the attribute features are prominent,and the decoupling ability between attributes is stronger.