TwinGAN: Twin Generative Adversarial Network for Chinese Landscape Painting Style Transfer

Abstract

Recently, style transfers have received considerable attention. However, most of these studies were suitable for Western paintings. In this paper, a deep learning method is proposed to imitate multiple styles of Chinese landscape paintings. Twin generative adversarial network style transfer was proposed based on the characteristics of Chinese landscape ink paintings. SketchGAN and renderGAN were performed using generative models based on generative adversarial networks. The SketchGAN involves determining the structure and simplifying the content of an input image. RenderGAN involves transferring the results of sketchGAN into the final stylized image. Moreover, a loss function was designed to maintain the shape of the input content image. Finally, the proposed TwinGAN was successfully used to imitate five styles of Chinese landscape ink paintings. This study also provided ablation studies and comparisons with previous works. The experimental results show that our algorithm synthesizes Chinese landscape stylized paintings that are higher in quality than those produced by previous algorithms.

Keywords

• Deep neural networks
• style transfer
• generative adversarial network (GAN)
• loss function
• Chinese landscape painting