Images Inpainting via Region Convolution and Feature Fusion

Image inpainting aims to fill damaged regions with reasonable structure, fine texture details, and semantically consistent content. However, existing deep learning algorithms mostly employ standard convolution architecture on damaged images, resulting in meaningless content such as color differences, blurring, and other artifacts. Aiming at the problem that the standard convolution methods cannot fully exploit the known regions to predict the damaged region features, this paper proposes an efficient region convolution (ERC) module to distinguish between known and damaged regions, and effectively prevent interference from invalid pixels through feature migration, additionally, the sampling regions are dynamically expanded to reconstruct the damaged regions more accurately. This article proposes a novel inpainting network that utilizes an encoder-decoder framework to restore images that are both semantically reasonable and visually realistic. Extensive experiments on two public datasets, Paris StreetView and CelebA, both qualitatively and quantitatively prove the strengths of the proposed approach.