Complex Manga Coloring Method Based on Improved Pix2Pix Model

In recent years, in the field of machine learning and computer vision, convolutional neural network (CNN) and conditional generative adversarial network (cGAN) have been widely used to generate realistic images, which can be used for coloring black and white comics, illustrations and line drafts. In this paper, we try to improve the pix2pix model based on the conditional GAN by using the multi-channel feature extraction to preserve the edge and grayscale information of complex images as much as possible. The proposed model uses multi-channel downsampling and fusion overlay technology to train and process complex illustrations, focusing on learning the color and contour lines of the picture respectively, to color complex black and white illustrations. For experimental needs, we prepared a training set with very different styles and complex pictures that we manually screened to broadens the range of applicable works. It can be seen from the results that this method retains the details of the original illustration, and performs color-changing processing on this basis, which has a good subjective evaluation and visual effect.