Region Hiding for Image Inpainting via Single-Image Training of U-Net
Chloe Martin-King, Mohamed Allali · 2019
In this paper we introduce the concept of region hiding as a clever way to inpaint an image using machine learning in which training relies solely on the damaged image itself. Region hiding encourages the network to learn the act of inpainting as it pertains to a specific image by hiding random intact regions and awarding accurate, coherent predictions of the hidden regions. This is done with respect to the entire image sans the truly missing or damaged areas. Once the network is trained to inpaint the intact hidden regions, it is employed to inpaint the truly missing regions of the damaged image. We show that using a U-net-like architecture with partial convolutional layers and transformation-based augmentation generates good image inpainting results even when the system is trained on a single image.