Image inpainting via multi-resolution network with Fourier convolutions

Nan Zhao, Li-Yong Shen, Ji-Wei Wang

Abstract: Image inpainting has been a trending topic among researchers in recent years, which aims to fill missing areas in images while maintaining visual and semantic consistency with the original image. Existing methods have their own limitations. While traditional convolution-based methods excel at filling most holes, they encounter difficulties when faced with large holes due to the limited receptive field. In contrast, Fourier convolutions-based methods exhibit effectiveness in filling large missing areas but have poor ability to restore local details. In this work, we propose the novel multi-resolution inpainting network with Fourier convolutions (MRFNet). Specifically, to better capture the correlation between semantic information in different regions, we first design the multi-scale large kernel attention (MSLKA) and multi-scale attention enhancement module (MSAE). Further, we use MSAE to construct our MRFNet. Ultimately, with the feature aggregation module proposed, our model enables high-fidelity image inpainting. Extensive experiments on two public datasets demonstrate the superior performance of the proposed model.