RFT: recurrent and Fourier transform for large hole image inpainting

Restoring damaged images using image restoration techniques is a simple task, but achieving more realistic effects faces huge challenges. We propose a two-stage image restoration model that innovatively combines the cyclic restoration mechanism with Fast Fourier Transform (FFT). By means of cyclic consistency constraints, the structural integrity of edge restoration is ensured, and then FFT is utilized to capture global information in the frequency domain, which significantly improves the accuracy and robustness of image restoration. In order to further improve the realism of the generated images, adversarial training is introduced in the second stage, and perceptual loss and color consistency loss are combined to ensure that the restored images are highly consistent with the original images in visual perception and color distribution. Experimental results show that our model has reached the state-of-the-art level on multiple public datasets, especially showing outstanding performance in edge restoration and color recovery tasks in complex scenarios.