A Proposal for Improvement of the AnimeGAN Based on CNN

In this paper, we propose a method to address the discoloration issue in the AnimeGAN model. Our method uses Convolutional Neural Networks (CNN) to sustain the chromatic integrity of the original image. We separate the image in the Lab color space, where the components of luminance (L), green-to-red variation (a), and blue-to-yellow gradient (b) exist independently. We use the L channel as the input for the AnimeGAN and color the grayscale image by predicting (a) and (b) channels with CNN. According to the experimental results, the discoloration problem of the AnimeGAN has been rectified, successfully preserving the color of the original image. Key Words: NST, GAN, AnimeGAN, CNN, Lab