Abstract
Japanese comics (called manga) are traditionally created in monochromeformat. In recent years, in addition to monochrome comics, full color comics, amore attractive medium, have appeared. Unfortunately, color comics requiremanual colorization, which incurs high labor costs. Although automaticcolorization methods have been recently proposed, most of them are designed forillustrations, not for comics. Unlike illustrations, since comics are composedof many consecutive images, the painting style must be consistent. To realizeconsistent colorization, we propose here a semi-automatic colorization methodbased on generative adversarial networks (GAN); the method learns the paintingstyle of a specific comic from small amount of training data. The proposedmethod takes a pair of a screen tone image and a flat colored image as input,and outputs a colorized image. Experiments show that the proposed methodachieves better performance than the existing alternatives.