ClsGAN: Selective Attribute Editing Model based on Classification Adversarial Network

Abstract Attribution editing has achieved remarkable progress in recent years owing to the encoder–decoder structure and generative adversarial network (GAN). However, it remains challenging generate high-quality images with accurate attribute transformation. Attacking these problems, work proposes a novel selective model based on classification (referred as ClsGAN) that shows good balance between transfer accuracy photo-realistic images. Considering are prone be affected by original due skip-connection structure, an upper convolution residual Tr-resnet) is presented selectively extract information from source image target label. In addition, further improve of generated images, classifier Atta-cls) introduced guide generator perspective through learning defects Experimental results CelebA demonstrate our ClsGAN performs favorably against state-of-the-art approaches quality accuracy. Moreover, ablation studies also designed verify great performance Tr-resnet Atta-cls.