Controllable style transfer of ink painting textures via contrastive learning

Yingjie Lin,Yi Zhang,Yan Cao,Kaijun Wu

DOI: https://doi.org/10.1109/IJCNN60899.2024.10651519

2024-06-30

Abstract:Style transfer technology has been successfully used in various visual art creation tasks and can generate various artistic styles. As a unique artistic product in the history of Chinese art, ink painting’s texture characteristics and color expression are crucial to expressing the ancient Chinese’s perception of personality and style. However, traditional texture transfer methods often difficult to achieve precise control of the texture of ink paintings. To help relevant practitioners better learn and draw lessons from the generated ink paintings, we proposed a controllable image texture style transfer model (CCTS) for ink paintings based on contrastive learning, which uses contrastive learning for style representation learning and texture-controlled migration. Our framework mainly consists of a texture control module, an image style transfer module, and a contrast learning encoding module. In the contrast learning module, we propose contrast enhancement of the input image as an enhanced style encoding to induce the network to generate style maps with more realistic colors. In addition, we propose a texture controller that performs convolution operations on encoded vectors of images of different resolutions of the same content, enabling users to control stroke styles and abstractions to generate different ink-painting textures. Extensive experiments based on the collected data sets show that in the task of ink painting texture art style transfer, this method has substantial improvements in both controllable texture transfer and ink painting color quality.

Computer Science,Art