Open AccessDiversity Regularized StarGAN for Multi-style Fonts Generation of Chinese Characters
Author(s) -
Jinshan Zeng,
Qi Chen,
Mingwen Wang
Publication year - 2021
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1880/1/012017
Subject(s) - font , computer science , calligraphy , generative grammar , chinese characters , generator (circuit theory) , focus (optics) , character (mathematics) , style (visual arts) , adversarial system , scalability , artificial intelligence , quality (philosophy) , exploit , painting , mathematics , power (physics) , art , philosophy , physics , geometry , literature , computer security , epistemology , quantum mechanics , database , optics , visual arts
The generation of stylish Chinese fonts plays a central role in many applications such as the design of art fonts and Chinese calligraphy generation. Most of existing methods focus on the generation of a single-style Chinese font, while few works focus on the multi-style font generation. In this paper, we exploit the star generative adversarial networks (StarGAN), a very popular generative adversarial networks (GAN) model recently developed in the literature, to realize the generation of multi-style Chinese fonts via a single model. Furthermore, in order to tackle the generation issue of Chinese characters having similar strokes for StarGAN, i.e., generating the same mode for these different but similar Chinese characters, we introduce a diversity regularizer such that the generator can generate high-quality characters with better diversity. A series of experiments are conducted on a handwritten Chinese character dataset called CASIA-HWDB1.1 and three standard printing font datasets to show the effectiveness of the proposed method. The experiment results show that the proposed method can effectively tackle the generation issue of Chinese characters having similar strokes in terms of the quality and diversity of generated results, via comparing to the baseline StarGAN, and is scalable to the multi-font generation via comparing to existing methods for the single-style font generation.