Into the Colorful World of Webtoons: Through the Lens of Neural Networks

The task of colorizing black and white images has previously been explored for natural images. In this paper we look at the task of colorization on a different domain: webtoons. To our knowledge this type of dataset hasn't been used before. Webtoons are usually produced in color thus they make a good dataset for analyzing different colorization models. Comics like webtoons also present some additional challenges over natural images, such as occlusion by speech bubbles and text. First we look at some of the previously introduced models' performance on this task and suggest modifications to address their problems. We propose a new model composed of two networks; one network generates sparse color information and a second network uses this generated color information as input to apply color to the whole image. These two networks are trained end-to-end. Our proposed model solves some of the problems observed with other architectures, resulting in better colorizations.