Scholarlink

<p>Document understanding in Indian scripts is challenging due to the non-linear grapheme composition, including conjunct consonants and spatially displaced vowel modifiers, which create complex two-dimensional character structures. To address these challenges, we introduce the IndicScript dataset, a newly curated and annotated dataset comprising Hindi, Bangla, and Odia document images for character segmentation and word-level information extraction. We further propose IndicCharGrid, a grapheme-aware 2D character grid representation that encodes each character at its precise spatial location, preserving both structural dependencies and semantic information. The grid representation is processed using a multi-scale Feature Pyramid Network (FPN) within an encoder–decoder architecture, enabling accurate character segmentation and entity extraction. We evaluate IndicCharGrid against baselines, including LayoutLMv3 (fine-tuned), Donut (OCR-free transformer), and a Transformer-based OCR pipeline. Experimental results shown in Fig. 1 demonstrate that IndicCharGrid consistently outperforms these approaches across all three languages, achieving F1-scores of 99.49%, 99.19%, and 98.06% on Hindi, Bangla, and Odia respectively improvements of up to 1.05 percentage points over the strongest baseline alongside Word Recognition Rates of 99.74%, 87.47%, and 90.95%, while also maintaining lower computational overhead.</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

IndicCharGrid: A Character Grid-Based Representation for Spatially-Aware Document Understanding in Indian Language Documents

  • 期刊论文

  • 发表日期：

    2026年5月29日

查看全文
  • Akkshita Trivedi,
  • Sandeep Khanna,
  • Santanu Chaudhury,
  • Gaurav Harit

摘要

Document understanding in Indian scripts is challenging due to the non-linear grapheme composition, including conjunct consonants and spatially displaced vowel modifiers, which create complex two-dimensional character structures. To address these challenges, we introduce the IndicScript dataset, a newly curated and annotated dataset comprising Hindi, Bangla, and Odia document images for character segmentation and word-level information extraction. We further propose IndicCharGrid, a grapheme-aware 2D character grid representation that encodes each character at its precise spatial location, preserving both structural dependencies and semantic information. The grid representation is processed using a multi-scale Feature Pyramid Network (FPN) within an encoder–decoder architecture, enabling accurate character segmentation and entity extraction. We evaluate IndicCharGrid against baselines, including LayoutLMv3 (fine-tuned), Donut (OCR-free transformer), and a Transformer-based OCR pipeline. Experimental results shown in Fig. 1 demonstrate that IndicCharGrid consistently outperforms these approaches across all three languages, achieving F1-scores of 99.49%, 99.19%, and 98.06% on Hindi, Bangla, and Odia respectively improvements of up to 1.05 percentage points over the strongest baseline alongside Word Recognition Rates of 99.74%, 87.47%, and 90.95%, while also maintaining lower computational overhead.