Deep Generative Inpainting with Comparative Sample Augmentation

Recent advancements in deep learning techniques such as Convolutional Neural Networks(CNN) and Generative Adversarial Networks(GAN) have achieved breakthroughs in the problem of semantic image inpainting, the task of reconstructing missing pixels in given images. While much more effective than conventional approaches, deep learning models require large datasets and great computational resources for training, and inpainting quality varies considerably when training data vary in size and diversity. To address these problems, we present in this paper a inpainting strategy of \textit{Comparative Sample Augmentation}, which enhances the quality of training set by filtering out irrelevant images and constructing additional images using information about the surrounding regions of the images to be inpainted. Experiments on multiple datasets demonstrate that our method extends the applicability of deep inpainting models to training sets with varying sizes, while maintaining inpainting quality as measured by qualitative and quantitative metrics for a large class of deep models, with little need for model-specific consideration.