Fast-GAN:User-Guided Sketch Colorization With Lightweight Generative Adversarial Networks

Sketch coloring refers to the generation of corresponding color images from given sketches or line drawings that lack sufficient semantic and texture information.With the development of artificial intelligence.image colorization has been rapidly developed and achieved gratifying results.It is widely used in education.film and other scenes.However.due to the differences between sketches and color images,current sketch coloring methods still have boundary overflow,watercolor artifacts and coloring efficiency problems.Therefore,it is of great significance to study the sketch coloring method based on user interaction.By analyzing the current research status and limitations of the current user-interactive sketch coloring method,this article mainly focuses on the two aspects of sketch coloring quality and coloring speed,and finally builds a lightweight framework called Fast-GAN.The main research contents are as follows:(1)The animation dataset and children’s drawing dataset are constructed.In order to train model,our constructs 30,000 color animation images and 15,000 color images of children’s drawings through python spider and constructs corresponding sketches by selecting appropriate outline extraction methods and manual screening.(2)A two-stage shading method based on texture constraints is proposed,which can overcome the problems of boundary overflow and watercolor artifacts in sketch shading.This method is inspired by the actual painting process.The first stage is the imitation stage.A Gray-Image Generator Network(GGN)is constructed.The corresponding gray-scale image is generated from the sketch through GGN,and a texture is designed.The feature loss function guarantees the quality of the grayscale image generated by the GGN network;the second stage of the method is the coloring stage,and a Color-Image Generator Network(CGN)is constructed,which is generated by the sketch and the first stage.The gray-scale image gets the final coloring result through the CGN network under user interaction.This method reasonably divides a difficult problem into two sub-problems.greatly improves the coloring quality,and can show excellent coloring effects on the line art drawings drawn by the artist.(3)A lightweight shading method.The method uses a lightweight feature extraction module to build a lightweight model,which makes the coloring speed reach real-time.At the same time,in order to overcome the problem of boundary overflow and watercolor artifacts in the sketch coloring process,based on the second research content,a double discriminator network architecture was designed to ensure the coloring quality.In addition,the algorithm can show good coloring effects on sketches drawn by people without a painting background.Ablation experiments and comparative experiments proved that our proposed coloring method can solve the problem of boundary overflow and watercolor artifacts in the process of sketch coloring.The verification of the manual line drawing shows that the method in this paper has a good generalization Ability.Finally,the coloring method is verified by a variety of evaluation indicators,and the experiment shows that the method in this paper can generate vivid color and high-quality color images.The proposed method promote the development and innovation of the research field of image coloring,and is of great significance to the development of children’s painting education.