| Arbitrary facial attribute editing refers to freely modifying multiple attributes of a facial image while ensuring that the other facial attributes remain unchanged,in order to obtain a fictitious new facial image that meets the user’s requirements.With the advent of Generative Adversarial Networks(GAN),arbitrary facial attribute editing based on GAN has become one of the hottest research directions in the field of deep learning.The gametheoretic approach of GANs has promoted the generation of more sophisticated and natural facial images in arbitrary facial attribute editing models.Nowadays,GAN-based arbitrary facial attribute editing technology has been widely applied in entertainment,medical aesthetics,and fake face detection,among other fields.However,the current arbitrary facial attribute editing models still face problems such as low image quality and inaccurate attribute editing.This paper explores the use of GANs for arbitrary facial attribute editing,with the following main research objectives:(1)Building upon the foundation of GAN,this paper addresses the issue of inaccurate facial attribute editing by designing a facial attribute editing model that incorporates the concept of latent code disentanglement.The current mainstream STGAN model uses the latent code concept of UNIT,which does not further segment into content latent code and style latent code,nor are they processed accordingly in the loss function.This can result in high coupling between the style and content parts of the generated image,making it difficult to accurately manipulate the desired facial attributes.This paper introduces the concept of latent code disentanglement on top of STGAN and redesigns the network architecture and loss function.In comparison experiments with mainstream models for facial attribute editing,the new model has higher accuracy in editing facial attributes and better qualitative and quantitative performance.(2)Building upon the foundation of GAN,this paper addresses the issue of lowquality generated facial images by designing a facial attribute editing model that combines pixel-level reconstruction loss and latent code reconstruction loss.Improving the model’s reconstruction ability can further enhance the image quality generated by the model.In order to improve the reconstruction ability,most models calculate the L1 or L2 loss between the reconstructed image and the input image at the pixel level,namely pixellevel reconstruction loss.These models minimize the pixel-level reconstruction loss to make the reconstructed image and input image more similar,thus improving the quality of the reconstructed image.However,using only pixel-level reconstruction loss cannot fully ensure the consistency between the reconstructed image and the input image at the latent code level.This paper simultaneously uses both pixel-level reconstruction loss and latent code reconstruction loss in the model,taking into account the consistency at both the pixel and latent code levels,further improving the model’s reconstruction ability.In comparison experiments with mainstream facial attribute editing models,the new model has a higher reconstruction ability and shows better qualitative and quantitative performance. |
