Chip-SAGAN: A Self-Attention Generative Adversarial Network for Chinese Ink Wash Painting Style Transfer

Abstract

The transfer of artistic style is a major and demanding task in computer vision. Compared with western paintings, Chinese ink wash paintings have unique characteristics that prevent existing methods from yielding satisfactory results, such as voids, brush strokes, and ink wash tone and diffusion. The main problems include: 1) the generator does not concentrate on key global features; 2) the generated paintings lose the color of the original content image; 3) the generated paintings do not have bright edges and smooth shading. In this paper, we propose Chip-SAGAN, an innovative approach to transforming real-world pictures into Chinese ink wash paintings, which is trained with unpaired photos and Chinese ink wash paintings. We introduce a self-attention module into the generator to capture global dependencies between features. Furthermore, we introduce an edge-promoting adversarial loss and a color reconstruction loss to ensure that the generated painting matches the content image’s edges and colors. The experimental results show that our method can transform real-world pictures into high-quality Chinese ink wash paintings, and surpass state-of-the-art algorithms.