An end-to-end model for chinese calligraphy generation

Peichi Zhou, Zipeng Zhao, Kang Zhang, Chen Li, Changbo Wang

Abstract: A Chinese calligraphy copybook usually has a limited number of Chinese characters, far from a whole set of characters needed for typesetting. Therefore, there is a need to develop complete sets of Chinese calligraphy libraries for well-known calligrapher styles. This paper proposes an end-to-end network for character generation based on specific calligraphy styles. Specifically, a style transfer network is designed to transfer the style of characters, and a content supplement network is designed to capture the details of stylish strokes. Our model can generate high-quality calligraphy images without manually annotating data. To verify the generated calligraphy styles, a new dataset is constructed for experimental comparison between our method and two other baseline methods. Moreover, a user study is conducted to evaluate our generated calligraphy from a visual perspective. When the experiment participants are asked to distinguish the real calligraphy from generated samples, the correct rate was 53.5%. The results show that the calligraphy styles generated by our model are almost indistinguishable from the original works.