Scholarlink

<p>Text detection in handwritten archival documents serves as a prerequisite for intelligent archive digitization and has witnessed substantial progress across diverse application domains. However, specialized detection methods and publicly available datasets for the Mongolian Historical Archive remain scarce. Several intrinsic characteristics of historical documents—including diverse and inconsistent textual corrections, varying levels of ink bleed-through, and irregular skew—pose significant challenges to accurate text detection. To tackle these issues, this paper presents a tailored text detection method for handwritten Mongolian historical documents. Specifically, to accommodate the varying correction styles in Mongolian manuscripts, we iteratively adjust the points within the initial candidate bounding boxes to ensure that the final contours fully enclose the corrected content. To address bleed-through artifacts resembling seal impressions, we incorporate an integrated attention mechanism to effectively distinguish textual regions from non-textual ones. Furthermore, we construct a dedicated dataset of Mongolian historical documents to facilitate future research. Experimental results demonstrate that the proposed network achieves accurate text detection on Mongolian archival documents, reaching an F-measure of 99.0%. The method also generalizes well to other historical datasets, with F-measures of 88.0% on DIVA-HisDB, 84.0% on ANDAR-TL-1&#xa0;K, and 73.5% on Handwritten Devanagari.</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

Contour thinning for text detection in Mongolian handwritten historical documents

  • 期刊论文

  • 发表日期：

    2026年2月4日

查看全文
  • Liuxu Ding,
  • Yuefeng Liu

摘要

Text detection in handwritten archival documents serves as a prerequisite for intelligent archive digitization and has witnessed substantial progress across diverse application domains. However, specialized detection methods and publicly available datasets for the Mongolian Historical Archive remain scarce. Several intrinsic characteristics of historical documents—including diverse and inconsistent textual corrections, varying levels of ink bleed-through, and irregular skew—pose significant challenges to accurate text detection. To tackle these issues, this paper presents a tailored text detection method for handwritten Mongolian historical documents. Specifically, to accommodate the varying correction styles in Mongolian manuscripts, we iteratively adjust the points within the initial candidate bounding boxes to ensure that the final contours fully enclose the corrected content. To address bleed-through artifacts resembling seal impressions, we incorporate an integrated attention mechanism to effectively distinguish textual regions from non-textual ones. Furthermore, we construct a dedicated dataset of Mongolian historical documents to facilitate future research. Experimental results demonstrate that the proposed network achieves accurate text detection on Mongolian archival documents, reaching an F-measure of 99.0%. The method also generalizes well to other historical datasets, with F-measures of 88.0% on DIVA-HisDB, 84.0% on ANDAR-TL-1 K, and 73.5% on Handwritten Devanagari.