Improving GAN-Based Calligraphy Character Generation using Graph Matching
Meng Li, Jian Wang, Yi Yang, Weixing Huang, Wenjuan Du · 2019
The Chinese character generation with specific style is the key of personal calligraphy font generation. In recent years, with the development of artificial intelligence, researchers have used deep learning to solve the calligraphy generation problem, which can automatically generate personal calligraphy fonts. However, the end-to-end approaches possibly generate wrong results in terms of glyph structure of Chinese character because of lack of constraints on glyph structures. This paper proposes a method that adds constraints of glyph structure to deep neural network in order to improve the correctness of generation. This paper represents the glyph structure of Chinese character with glyph nodes extracted by object detection method. We compute the glyph structure loss between inputs and the generated results by the deep learning-based graph matching method. Experiments show that our method significantly increases the accuracy of Chinese character generation. For showing our method more clearly, we use the print as the original style samples in this paper.