LocalDoc/azerbaijani-htr-benchmark

阿塞拜疆语手写OCR基准数据集是一个用于手写文本识别（HTR）的手动标注基准，针对阿塞拜疆语拉丁字母手写文本。数据集包含真实世界扫描的手写阿塞拜疆语页面，使用Label Studio进行标注，标注内容包括旋转边界框和精确转录。数据集以两种并行视图提供：lines（行级识别）包含裁剪和去歪斜的单行图像及其转录，共442行，来自29个源页面，平均每页15.2行；pages（全页结构化标注）包含原始扫描图像和所有行的轴对齐边界框及转录。数据集构建过程包括收集真实手写页面、手动标注、处理旋转区域（行级视图进行去歪斜，全页视图转换边界框），并可选将页面图像缩放至最大2200像素以减小文件大小。转录为字面转录，保留手写错误。数据遵循CC BY 4.0许可。

A manually annotated benchmark for handwritten text recognition (HTR) on Azerbaijani Latin script. Real-world scanned pages annotated in Label Studio with rotated bounding boxes and exact transcriptions. Provided in two parallel views: lines for line-level recognition with cropped and deskewed single text line images paired with transcriptions (total 442 lines from 29 source pages, avg 15.2 lines per page), and pages for full-page structured annotations with original scanned pages and axis-aligned bounding boxes and transcriptions of all lines. Built from collected Azerbaijani handwritten pages, manually annotated, with rotated regions processed (deskewed for lines, converted to axis-aligned boxes for pages), and page images optionally downscaled to max 2200px. Transcriptions are literal, preserving handwriting errors. Licensed under CC BY 4.0.