A multi-level feature integration network for image inpainting

Deep learning-based methods have shown great potential in image inpainting, especially when dealing with large missing regions. However, the inpainted results often suffer from blurring, and improper textures can be created without an understanding of semantic information. In order to extract more features from the known regions, we propose a multi-level feature integration (MFI) network for image inpainting. We complete hole regions by two generators. For each generator, we use the MFI network to fill the hole region with multi-level skip connections. With multi-level feature integration, the network gains more knowledge about the global semantic structures and local fine details. Moreover, instead of a deconvolution layer or an interpolation algorithm, we adopt a sub-pixel layer to up-sample feature maps and produce more coherent results. We use PatchGAN to support the refinement generator network to produce more discriminative detail. Our experiments done with the Paris StreetView, CelebA-HQ and Places2 datasets demonstrate the effectiveness of our MFI network for producing visually pleasing results with semantically ordered textures.