Style-independent radical sequence learning for zero-shot recognition of Small Seal script

Abstract

Small Seal script (Chinese called XiaoZhuan), as the earliest archaic form of standardized Chinese script, is the predecessor of modern Chinese characters. However, the Small Seal character recognition remains a challenging task, due to various un-/semi-structured pictographic glyphs and writing styles. This paper proposes a style-independent pictographic radical decomposition for the zero-shot recognition of Small Seal script, by taking advantage of the inherent consistency of pictographic representations between the Small Seal script and the Traditional Chinese script (Chinese called Fanti). Specifically, we design a feature-level collaboration framework of two tasks. One is the XiaoZhuan-to-Fanti translation task, which employs a generative adversarial network (GAN) based dual-learning mechanism to learn style-independent and consistent pictographic feature representations from different styles of Small Seal and their corresponding Traditional Chinese characters. The other is a transformer-based pictographic radical sequence learning from the pictographic feature representations. Experiments demonstrate that our model has satisfactory recognition ability to various styles of Small Seal scripts, especially for the zero-shot recognition of those with unknown glyphs and unseen styles. The code is available at https://github.com/windyz77/SmallSealRecon.