Cross-Domain Learning for Reference-Based Sketch Colorization with Structural and Colorific Strategy.

This paper aims to tackle the colorization task of sketch image given an already-colored reference image. Sketch colorization is a thorny task for computer vision since neither grayscale values nor semantic information exists in sketch images. To address this, We propose to jointly train the domain alignment network with a simple adversarial strategy, that we term the structural and colorific conditions, to learn the semantical correspondence between information-scarce sketch and the given instructive reference. Specifically, the inputs from distinct domains will be aligned to an embedding space where the semantical correspondence is established, then, the generator will reconstruct the sketch image according to the established correspondence. We demonstrate the effectiveness of our proposed method in sketch colorization tasks via quantitative and qualitative evaluation against existing approaches in terms of image quality as well as style relevance.