"Handwritten Devanagari COnjunct COnsonants (HDCOCO) Dataset"

"Handwritten optical character recognition (OCR) for Devanagari has advanced rapidly, yet conjunct consonants remain a major bottleneck due to complex ligatures and diverse writing styles. We introduce Handwritten Devanagari COnjunct COnsonants (HDCOCO), a curated dataset of 50,020 handwritten samples spanning 41 conjunct classes, collected across multiple institutions and public venues to capture wide demographic and stylistic variation. We validate HDCOCO through embedding-space analyses, PCA\/t-SNE cluster structure, centroid-distance difficulty scoring, and stratified train--test splits matched by difficulty, confirming class separability and fair evaluation conditions. Our contribution provides a solid foundation for state-of-the-art Devanagari OCR on real-world handwriting."

针对天城文（Devanagari）的手写光学字符识别（OCR）技术近年来发展迅猛，但受限于复杂的连字结构与多样的书写风格，辅音连缀仍是该领域的核心瓶颈。我们提出了手写天城文辅音连缀数据集（Handwritten Devanagari Conjunct Consonants, HDCOCO）：该数据集经精心筛选整理，包含50020条手写样本，涵盖41类辅音连缀，采集自多所科研机构与公共场景，以覆盖广泛的人群分布与书写风格差异。我们通过嵌入空间分析、主成分分析（PCA）/t分布邻域嵌入（t-SNE）聚类结构、质心距离难度评分，以及按难度匹配的分层训练-测试划分对HDCOCO进行了有效性验证，证实了该数据集的类别可分性与公平的评估条件。本研究成果可为面向真实手写场景的顶尖天城文OCR技术提供坚实的研究基础。