Sketch to Chinese paintings: A three-stage progressive generation network via enhancing sketch

Abstract

With the proposal and wide application of Generative Adversarial Networks (GAN), sketch-based image generation has gradually become a research hotspot. Because of its unique artistic characteristics, Chinese painting has attracted more and more people to engage in research in the field of sketch-based Chinese painting. Most existing researches on sketch generation of Chinese paintings tend to extract edge maps from mature Chinese paintings and train generative models. When edge maps are extracted from sketches with sparse lines as input for generation, the quality of the generated Chinese painting is poor. This paper proposes a three-stage progressive Chinese painting generation network based on sketch. By the reduction and enhancement networks, our model converts the input sketch into types of sketches with different line richness. Each stage is used to learn to generate different Chinese painting information, realizing the progressive generation of Chinese painting through three connected generation networks. The experimental results show that our model can generate better-quality Chinese paintings and perform better in generating Chinese paintings from sketches.