Hui Zheng ; Li Jie ; Wang Xiumei ; Gao Xinbo

Abstract

Image inpainting techniques have shown promising improvement with theassistance of generative adversarial networks (GANs) recently. However, most ofthem often suffered from completed results with unreasonable structure orblurriness. To mitigate this problem, in this paper, we present a one-stagemodel that utilizes dense combinations of dilated convolutions to obtain largerand more effective receptive fields. Benefited from the property of thisnetwork, we can more easily recover large regions in an incomplete image. Tobetter train this efficient generator, except for frequently-used VGG featurematching loss, we design a novel self-guided regression loss for concentratingon uncertain areas and enhancing the semantic details. Besides, we devise ageometrical alignment constraint item to compensate for the pixel-baseddistance between prediction features and ground-truth ones. We also employ adiscriminator with local and global branches to ensure local-global contentsconsistency. To further improve the quality of generated images, discriminatorfeature matching on the local branch is introduced, which dynamically minimizesthe similarity of intermediate features between synthetic and ground-truthpatches. Extensive experiments on several public datasets demonstrate that ourapproach outperforms current state-of-the-art methods. Code is available athttps://github.com/Zheng222/DMFN.