Scientific Data · November 11, 2026
A Paired Printed-Handwritten Arabic Line Dataset for Handwriting Recognition and Cross-Modal Retrieval
Mahmoud SalahEldin Kasem, Abdelrahman Abdallah, Mohamed Mahmoud, Mahmoud Abdalla, Ibrahim Abdelhalim, Islam Taj-Eddin
Abstract
Automated recognition of Arabic handwriting lags behind Latin-script systems, partly due to the scarcity of openly available datasets that pair printed ground-truth (GT) lines with authentic handwritten (HW) counterparts. We introduce SAQR (from the Arabic word saqr , meaning “falcon”), a moderately sized, regionally focused dataset of 2,231 paired GT–HW Arabic text line images collected from 323 adult undergraduate writers aged 18–22 at the Faculty of Computers and Information, Assiut University, Egypt. Each pair is accompanied by a manually verified Arabic transcription, a pseudonymous writer identifier, a recorded gender category, a form-template identifier, and a writer-independent train/validation/test split. All 2,231 records were manually inspected for transcription correctness, GT–HW correspondence, crop completeness, and metadata consistency. The corpus contains 24,974 whitespace-delimited word tokens, 3,996 observed word types, 111,221 characters across 47 Arabic character forms, and 1,296 distinct literal transcription strings (1,276 after Unicode and whitespace normalization). We establish initial reference baselines across three tasks: handwritten text recognition, exploratory gender classification from handwriting, and cross-modal GT–HW line retrieval. SAQR provides an openly available paired resource for handwriting recognition, demographic analysis, and cross-modal retrieval. It is intended as a moderately sized, regionally focused testbed rather than as a broadly representative sample of Arabic handwriting, distributed under a CC BY 4.0 license, with source attribution documented in the public repository.
More by Abdelrahman Abdallah
- A multilingual Arabic and English receipt dataset for text detection, recognition and information extractionZenodo · 2026
- UIBK-Embedding: Effective Text Retrieval by Unifying Dense and Late-Interaction RetrievalarXiv · 2026
- MemRevise: A Self-Revising Memory for Conversational SearcharXiv · 2026
- MM-BRIGHT: A Multi-Task Multimodal Benchmark for Reasoning-Intensive RetrievalACM SIGKDD Conference on Knowledge Discovery and Data Mining · 2026 · Citations: 12
