Attentional coarse-and-fine generative adversarial networks for image inpainting

Abstract Deep learning based image inpainting methods have recently made significant improvements, which can now generate visually realistic and semantically plausible pixels for large missing regions. However, blurry textures and distorted edges are the two main problems of these methods. In nature, they are caused by the ineffectiveness of utilizing features. Thus, we propose a coarse-and-fine structure, which can extract coarse features with a larger receptive field to guarantee the accuracy of global semantic information, and can simultaneously extract fine features with a smaller receptive field at multiple levels to serve as a supplement in a parallel manner. To avoid the artifacts caused by the missing part of image, we flexibly use the gating mechanism and propose an interleaved gated residual block (IGRB) to encourage the useful information flow through the neural network. Moreover, we propose a channel and spatial attention block (CSAB) to alleviate the influence of redundant information and better model the long-range dependency between different regions in the image. Extensive experiments on faces, natural objects and scenes demonstrate that our method outperforms the existing state-of-the-art methods.