ENC-PSL/MEDUSA_Synthetic_Lines

该数据集包含50万张合成行图像，为MEDUSA训练课程的Silver阶段生成。它由École nationale des chartes – PSL作为MEDUSA项目的一部分制作，用于多语言中世纪手写文本识别（HTR）。数据集旨在通过将文本资源渲染为逼真的合成手稿行图像，为MEDUSA模型提供语言和脚本级别的先验知识，特别是针对那些在现有图像-文本HTR语料库中代表性不足的中世纪语言（如日耳曼语、凯尔特语和斯拉夫语）。生成管道包括片段采样、渲染、墨水模拟、背景合成、扫描退化和文档效果等六个阶段，输出为JPEG图像和对应的JSONL元数据。数据集结构按语言组织，包含JPEG/ALTO XML文件对，适用于DocWorkflow工具。This dataset contains 500,000 synthetic line images generated for the Silver stage of the MEDUSA training curriculum. It was produced at the École nationale des chartes – PSL as part of the MEDUSA project for multilingual medieval handwritten text recognition (HTR). The dataset operationalises text-only resources by rendering them as plausible synthetic manuscript line images, providing the MEDUSA model with lexical and script-level priors for languages that are underrepresented in existing image–text HTR corpora, particularly Germanic, Celtic, and Slavic medieval languages. The generation pipeline involves six stages: snippet sampling, rendering, ink simulation, background composition, scan degradation, and document effects, outputting JPEG images with corresponding JSONL metadata. The dataset is structured by language with paired JPEG/ALTO XML files, ready for use with DocWorkflow.