Globally Consistent Image Inpainting based on WGAN-GP Network Optimization

There have been many methods applied to image inpainting. Although these algorithms can roughly produce visually plausible image structure and texture, they also create a lot of chaotic structural information and blurry texture details, resulting in inconsistencies with the surrounding content area. This paper proposes a globally consistent image inpainting network with a nonlocal module based on WGAN-GP optimization. It can make the network obtain the relevant information on long-distance dependence without superimposing network layers. And it is also able to prevent the limitations such as inefficient calculation and complex optimization caused by the local operation of the convolutional neural network. Thus making full use of the surrounding information of the area to be repaired will improve the semantic and structural consistency of generating predictions with the entire background area. Experiments with this model are conducted on a Places2 dataset, and the results prove that our method was superior to ordinary convolutional neural networks.