Scholarlink

<p>Layout analysis is central to document image processing, supporting tasks such as OCR, digital archiving, and cultural heritage preservation. Yet, annotating Arabic manuscripts remains challenging due to script-specific complexities, structural variability, and frequent degradation. This paper presents a semi-automatic annotation framework tailored for Arabic document images. The system integrates automatic preprocessing methods—including connected component extraction, sub-word and line segmentation, and figure detection—with a user-friendly interface for interactive refinement. This hybrid approach reduces manual effort while ensuring high annotation accuracy. Evaluation on a corpus of 1,000 Arabic pages from multiple benchmark datasets shows superior performance compared to GEDI, Aletheia, Transkribus, eScriptorium, and dhSegment, achieving an IoU of 0.8793 and an F1-score of 0.8902 while reducing annotation time by 60% relative to GEDI. Tests confirm its effectiveness, with annotators reporting less correction effort and higher satisfaction. The framework also produced ground truth for the SDADDS-Guelma dataset, highlighting its practical utility.</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

Semi-automatic framework for layout annotation of arabic document images

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Abderrahmane Kefali,
  • Ismail Bouacha,
  • Ryene Benatti,
  • Nor Elhouda Chaib Rassou

摘要

Layout analysis is central to document image processing, supporting tasks such as OCR, digital archiving, and cultural heritage preservation. Yet, annotating Arabic manuscripts remains challenging due to script-specific complexities, structural variability, and frequent degradation. This paper presents a semi-automatic annotation framework tailored for Arabic document images. The system integrates automatic preprocessing methods—including connected component extraction, sub-word and line segmentation, and figure detection—with a user-friendly interface for interactive refinement. This hybrid approach reduces manual effort while ensuring high annotation accuracy. Evaluation on a corpus of 1,000 Arabic pages from multiple benchmark datasets shows superior performance compared to GEDI, Aletheia, Transkribus, eScriptorium, and dhSegment, achieving an IoU of 0.8793 and an F1-score of 0.8902 while reducing annotation time by 60% relative to GEDI. Tests confirm its effectiveness, with annotators reporting less correction effort and higher satisfaction. The framework also produced ground truth for the SDADDS-Guelma dataset, highlighting its practical utility.