Convolutional Neural Network Style Transfer towards Chinese Paintings

Abstract

Chinese painting is one of the most important cultural heritages. However, creating a Chinese painting usually requires specific skills, patience, and years of professional training. Moreover, most existing style transfer methods mainly focus on photograph or western painting, and there are intrinsic differences between Chinese and western paintings. To this end, we propose a novel algorithm of Chinese painting style transfer (CPST) towards transferring with unique ink and wash characteristics, which automatically generates Chinese paintings with machine learning technology. In this paper, firstly, comparing Chinese paintings with western works, we set up four key restrictions in the style transfer, i.e., with special considerations of typical ink and wash features, including brush stroke, space reservation, ink tone with diffusion and yellowing. In order to incorporate these restrictions into the transferring convolutional neural networks (CNN), we separate different layers of the CNN into layers of style and content, so as to faithfully reserve the style of the reference image. Secondly, as Chinese paintings can be divided into fine brushwork and freehand brushwork, to cope with diverse painting skills, we devise different strategies to transfer not only the ink tone but also the painting skills to the target images. Experiments show that taking the aesthetic characteristics of Chinese painting style as reference, our results are more visually satisfactory and successfully overcome spillovers.