| In painting,coloring is one of the most time-consuming steps.To create an impressive and expressive painting requires not only a good color combination but also the appropriate use of light and shade and texture of the pattern.The traditional gray-scale image coloring method uses the continuity feature of the gray-scale area contained in the gray-scale image to divide the area with the same gray-scale information in the image to achieve the purpose of coloring.However,the black-and-white line draft image is mainly composed of lines,which lacks gray areas with distinct level differences and continuity.Therefore,the method of colorizing grayscale images cannot effectively colorize line art images.Under the background that deep learning technology driven by big data is becoming more and more mature,the automatic coloring technology of comic line art images has brought new opportunities for the rapid and high-quality production of animation.To solve the problem that the color rendering process of comic line art is too complicated,this article designs and implements the automatic rendering method and user-guided auxiliary coloring method of comic line art images based on conditional generation adversarial network to simplify the coloring process.Main tasks as follows:(1)An automatic coloring method for comic line art images based on conditional generation adversarial network is proposed.In order to solve the problem of poor automatic model coloring caused by the difference between the extracted line art image and the real line art image,the conditional generation adversarial network combined with pre-trained local feature extraction networks to effectively improve the model’s effect on real generalization ability of manuscript images.And by using Res Ne Xt module in the generator and discriminator network to enhance the learning ability of the model.Compared with Paints Chainer’s Tanpopo,Satsuki and Canna automatic coloring methods,this method can achieve a smaller Frechet Inception distance score and achieve better coloring results.(2)A method for assisted coloring of manga line art images based on user color strokes is proposed.In order to solve the problem that the automatic coloring method’s coloring effect on the line art image is too single and cannot meet the user’s diversity requirements,the network structure of the generator is changed.During the training stage,simulated image of user’s color strokes is used as the conditional input of the generator,which guides the model to produce different coloring effect.Compared with the user-guided coloring methods of Paints Chainer’s Tanpopo,Satsuki and Canna,this method achieves a higher mean opinion score and makes it easier for users to perform secondary coloring.(3)A method for assisting color rendering of line art images based on reference color images is presented.Aiming at the problem that the existing coloring methods do not take advantage of the existing color image,the pre-trained Visual Geometry Group network is used to extract the color distribution characteristics of the reference color image,and then guide the model to color.Secondly,the SE-Res Ne Xt sub-module is designed to automatically learn the feature relationship between different channels,so that the model makes better use of the color distribution features and improves the coloring effect of the line draft image.This method can generate colorized results in the same color style as the reference color image,and achieve good coloring results. |
