20.FaRE: A Feature-aware Radical Encoding Strategy for Zero-shot Chinese Character Recognition

Due to the complexity of glyphs and the vast vocabulary, zero-shot Chinese character recognition (ZSCCR) remains a prominent research topic. A mainstream approach involves radical-based character decomposition. However, existing methods typically employ random encoding for each radical post-decomposition, leading to potential topology distortions in the radical encoding and glyph spaces. To address these issues, we propose a novel Feature-aware Radical Encoding (FaRE) strategy that incorporates visual feature clues into radical encodings to generate feature-aware representations. Initially, we create radical images by rendering TTF files and then apply a pre-trained feature extractor to obtain the feature representation of each radical. Finally, projection and binarization operations are performed to produce compact and efficient radical encodings. Extensive experiments on the public benchmark ICDAR2013 demonstrate that the proposed FaRE significantly enhances the state-of-the-art ZSCCR performance. Additionally, abundant ablation studies are conducted to validate the effectiveness of the proposed FaRE.

FaRE: A Feature-aware Radical Encoding Strategy for Zero-shot Chinese Character Recognition, Zhan Hongjian and Li Yangfu and Xiong Yu-jie and Lu Yue,Proceedings of the Asian Conference on Computer Vision (ACCV),2024,390-401.