Creating New Chinese Fonts based on Manifold Learning and Adversarial Networks

Abstract

The design of fonts, especially Chinese fonts, is known as a tough task that requires considerable time and professional skills. In this paper, we propose a method to easily generate Chinese font libraries in new styles based on manifold learning and adversarial networks. Starting from a number of existing fonts that cover various styles, we firstly use convolutional neural networks to obtain the representation features of these fonts, and then build a font manifold via non-linear mapping. Using the font manifold, we can interpolate and move between those existing fonts to get new font features, which are then fed into a generative network learned via adversarial training to generate the whole new font libraries. Experimental results demonstrate that high-quality Chinese fonts in various new styles against existing ones can be efficiently generated using our method.