MRS-Net: an image inpainting algorithm with multi-scale residual attention fusion

Image repair is to repair partially damaged images. At present, when repairing the images with arbitrary missing shape and the images with large defect area, the current methods have some problems, such as fuzzy image repair and differences in the repaired joints. Therefore, this paper proposes an image repair model of MRS-Net(Multiscale Residual Squeeze-and-congestion Networks) by using the generated countermeasure network to repair the images with arbitrary missing shape and large defect area. The generator adds a residual attention module in the connection layer between the encoder and the decoder to improve the repair ability of the model, and generates a network through multi-scale joint feedback against loss, reconstruction loss, perception loss, style loss and total variation loss, so as to ensure the visual consistency between the repair boundary and the surrounding real image. At the same time, the binary cross entropy loss feedback discriminant network is used. The proposed model is trained and tested on the dataset CelebA and Oxford buildings. Experiments show that the proposed model can effectively extract the missing information, Meanwhile, the repair results have natural transition boundaries and clear details. MRS-Net can improve SSIM(structural similarity) index by about 2% - 5% and PSNR(peak signal-to-noise ratio)index by about 1-3. The FID(Frechet Inception Distance score) straight index is reduced by 2-8, and the evaluation indexes are improved.when the missing area accounts for 5% - 10%, 11% - 20%, 21% - 30%, 31% - 40% and 41% - 50%,MRS-Net has better image restoration effect, which can reflect that MRS-Net has better robustness.The proposed image restoration method can repair faces, buildings and other scenes. At the same time, it can repair defects in different shapes and areas.