RRecT: Chinese Text Recognition with Radical-Enhanced Recognition Transformer

Xinrui Deng, Zheng Huang, Kefan Ma, Kai Chen, Jie Guo, Weidong Qiu

Abstract: Text recognition has attracted continuous attention in recent years and reaches increasingly high performance on English datasets. Nevertheless, little attention is emphatically paid to Chinese Text Recognition (CTR), leading to a barely satisfying accuracy on Chinese datasets. Due to the complex glyph structure and large size of character set, CTR is more challenging and requires a powerful capacity of feature extraction. In this paper, we propose a novel network for CTR named Radical-enhanced Recognition Transformer (RRecT). It firstly introduces a customized Recognition Transformer (RecT) to extract multi-grained features, then exploits radical decomposition as an auxiliary supervision signal and enhances character representation with radical information by Radical Prediction Module (RPM) and Radical-Character Fusion Module (RCFM). Thus, final feature contains both character-level and fused radical-level information. The experimental results show that RRecT outperforms the state-of-the-art methods by a margin of 1.4% on Scene dataset, 1.8% on Document dataset and reaches a competitive performance on Web and Handwritten dataset. Moreover, RRecT requires much less computation cost and is a lightweight and effective model.