An Improved Method for Semantic Image Inpainting with GANs: Progressive Inpainting

Semantic image inpainting is getting more and more attention due to its increasing usage. Existing methods make inference based on either local data or external information. Generating Adversarial Networks, as a research focus in recent years, has been proven to be useful in inpainting work. One of the most representative is the deep-generative-model-based approach, which use undamaged images for training and repair the corrupted image with the trained networks. However, this method is too dependent on the training process, easily resulting in the completed image blurry in details. In this paper, we propose an improved method named progressive inpainting. With the trained networks, we use back-propagation to find the most appropriate input distribution and use the generator to repair the corrupted image. Instead of repairing the image in one step, we take a pyramid strategy from a low-resolution image to higher one, with the purpose of getting a clear completed image and reducing the reliance on the training process. The advantage of progressive inpainting is that we can predict the general distribution of the corrupted image and then gradually refine the details. Experiment results on two datasets show that our method successfully reconstructs the image and outperforms most existing methods.