CTRF-GAN: A Generative Network for Image Inpainting Using CSWin Transformer and Fast Fourier Convolution Residual

Image inpainting is an important task in computer vision, aiming at restoring missing or damaged areas in an image. The existing methods have certain problems such as texture blur and structure distortion, especially when the degenerated images exhibit complex structures and scenes. To address these issues, this paper proposes an image inpainting algorithm (called CTRF-GAN) based on the generative adversarial network (GAN) with CSWin Transformer module and Fast Fourier convolution residual module. The modified GAN framework replaces ordinary convolution in the encoder and decoder with gated convolution to reduce the training loss. The CSWin Transformer module is introduced into the generator to enhance the global dependency and to increase the receptive field. A fast Fourier convolution residual module (Res-AFFC) is proposed to extract high-frequency texture information. Finally, the HSV loss function is integrated to ensure that the color of the inpainting areas is consistent with the original image. Extensive experiments are conducted on the CelebA and Paris Street View datasets, verifying the superiority of the proposed algorithm.