Scholarlink

<p>The Tibetan script serves as a fundamental vehicle of Tibetan culture. Developing robust digital recognition methods is essential for preserving this cultural heritage. Existing approaches, primarily designed for single fonts, lack the capability for unified modeling of diverse writing variants such as Uchen and Umê. This paper proposes an ABINet-Irrelevant-Writing-Style (ABINet-IWS) model. The model introduces a Multi-strategy Masking Mechanism (MSMM) to enhance reasoning under incomplete information conditions and a Decoupling of Text Styles (DTS) module to isolate font-specific attributes from textual content representations. A Multi-font Tibetan Recognition Dataset (MFTR dataset) is also constructed, containing 16,735 images across five major fonts. Experiments on the MFTR dataset show that ABINet-IWS achieves character and word accuracy rates of 96.8% and 83.7%, substantially outperforming the baseline model. Evaluations on multiple public benchmarks further confirm the superior generalization capability of the proposed model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-font Tibetan script recognition: a unified modeling approach with masking mechanism and style decoupling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月25日

查看全文
  • Zhenjie Wu,
  • Weilan Wang,
  • Kunhao Xiao,
  • Yusheng Hao,
  • Leer Mao,
  • Xiaoping Shi,
  • Dazhi Yang,
  • Yajuan Lu

摘要

The Tibetan script serves as a fundamental vehicle of Tibetan culture. Developing robust digital recognition methods is essential for preserving this cultural heritage. Existing approaches, primarily designed for single fonts, lack the capability for unified modeling of diverse writing variants such as Uchen and Umê. This paper proposes an ABINet-Irrelevant-Writing-Style (ABINet-IWS) model. The model introduces a Multi-strategy Masking Mechanism (MSMM) to enhance reasoning under incomplete information conditions and a Decoupling of Text Styles (DTS) module to isolate font-specific attributes from textual content representations. A Multi-font Tibetan Recognition Dataset (MFTR dataset) is also constructed, containing 16,735 images across five major fonts. Experiments on the MFTR dataset show that ABINet-IWS achieves character and word accuracy rates of 96.8% and 83.7%, substantially outperforming the baseline model. Evaluations on multiple public benchmarks further confirm the superior generalization capability of the proposed model.