Chinese Font Generation Algorithm with Redundant Feature Suppression

Xiaoqi Guo, Cunrui Wang, Guodong Zhan, Juxiang Zhou

Abstract: Automated generation of Chinese character fonts can significantly enhance font library design efficiency and improve industrial productivity. Current algorithms for generating Chinese fonts often transmit redundant features from reference graphics into the model structure, alongside spatial and channel dimensions, resulting in pronounced redundancy in the generated outputs. This paper proposes a method that suppresses redundant features from both spatial and channel dimensions, reconstructs spatial features, and employs residual quantization to extract detailed font style characteristics. By mitigating redundant features, the overall stylistic expression of Chinese characters is reinforced, leading to the generation of high-quality font libraries. Experiments conducted on a dataset comprising 30 Chinese font libraries demonstrate the effectiveness of this method, with improvements observed in font feature expression metrics compared to CycleGAN, Zi $\mathrm{i}2\text{zi}$, MX-Font, and DG-Font.