kurdish-word-dataset (ckb-HWord): A Multi-Writer Offline Handwritten Central Kurdish Word Corpus
Description
The ckb-HWord (Kurdish Word Dataset) repository is a comprehensive, multi-writer offline handwritten Central Kurdish (Sorani) word corpus designed to advance optical character recognition (OCR), handwritten text recognition (HTR), and document intelligence models for low-resource extended Arabic scripts.The dataset contains 47,340 cropped word images covering 9,472 unique lexical items derived from an underlying sampling lexicon of 10,001 target words partitioned across 28 initial-letter classes. A total of 953 completed physical A4 collection sheets filled out by diverse demographic participants using various writing instruments were digitized at 600 DPI using a high-resolution flatbed optical scanner into 30 master PDF scan files. An automated Python processing pipeline (utilizing OpenCV, PyMuPDF, Pillow, and PyTesseract) executed grid alignment, table cell cropping, vertical scanner streak inpainting, and demographic checkbox extraction. Quality control protocols excluded 57 unextracted or unverified forms.The repository provides two parallel image sets, exhaustive ground-truth metadata, master scans, and end-to-end execution scripts:images/ (Primary RGB Set): 47,340 full-color 24-bit RGB PNG crops at native 600 DPI (typical size: 793 x 410 px), organized into 953 form-level subfolders.images_grayscale/ (Derived Greyscale Set): 47,340 single-channel 8-bit greyscale PNG crops vertically padded with white margins (pixel value = 255) to a uniform 512 px height while retaining native horizontal crop widths (typical size: 793 x 512 px).Structured Ground-Truth Metadata: forms_metadata.json (writer demographics, form IDs, crop coordinates), corpus_index.json (28 letter classes, Unicode codepoints, lexical frequencies), and manifest.csv (relative image paths mapped to ground-truth labels).Master Scans & Utilities: scanned_papers/ (30 high-resolution PDF scans), data/ (source lexical files), and Python scripts for extraction, cleaning, greyscale normalization, and verification.
Files
Institutions
- University of HalabjaSulaymaniyah, Halabja