Research map / Image generation / Text-to-image diffusion
Muharaf: Manuscripts of Handwritten Arabic Dataset for Cursive Text Recognition
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Muharaf, a dataset of 1,644 expert-transcribed historic handwritten Arabic manuscript pages (1,216 public) with line polygons, page elements, and a CNN-based SFR baseline for full-page HTR.
Problem
Handwritten text recognition for cursive historical Arabic manuscripts is hampered by a scarcity of large, publicly available, fully transcribed datasets of genuine (non-scribed) handwriting.
Approach
A large, expert-transcribed collection of original historic Arabic handwritten pages with polygonal text-line and page-element annotations, built with a semi-automatic loop in which an SFR model's predictions are manually corrected and used to retrain it.
Key results
- SFR trained on Muharaf over three random splits achieves page-level HTR results with CER below 20% after 500 training images (trial 5 onward).
- Combining Muharaf with RASAM and RASM improved SFR cross-dataset CER on KHATT by roughly 16%.
- TrOCR outperforms SFR on IAM but has 153.1M parameters versus SFR's 18M-parameter HW network.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train SFR with more collected Muharaf data | SFR · Muharaf | CER: below 20% | Helped |
| Combine Muharaf, RASAM, and RASM training data | SFR · KHATT unique-paragraph test lines | CER: ~16% | Helped |
| Use TrOCR instead of SFR | HTR on IAM · IAM | CER | Helped |
Surprising findings
- Models trained on Muharaf alone or on RASAM+RASM alone generalize poorly to KHATT, with only combined training giving a notable gain.
Limitations
- Only 1,216 of 1,644 pages (428 restricted) are publicly released under a proprietary license for the rest.
- Many line polygons are automatically generated and not tight, so annotations overlap.
- Writer identities and timelines are incomplete for many documents.
Experimental setup
Datasets: Muharaf IAM KHATT RASAM RASM
Benchmarks: Muharaf IAM KHATT
Metrics: character error rate word error rate
Baselines: TrOCR SFR on KHATT cross-dataset
Builds on: Start, Follow, Read PAGE-XML
Closest papers by meaning
- On the Generalization of Handwritten Text Recognition Models (CVPR 2025 · similarity 0.84)
- MetaWriter: Personalized Handwritten Text Recognition Using Meta-Learned Prompt Tuning (CVPR 2025 · similarity 0.84)
- Uni-MuMER: Unified Multi-Task Fine-Tuning of Vision-Language Model for Handwritten Mathematical Expression Recognition (NeurIPS 2025 · similarity 0.83)
- General Detection-based Text Line Recognition (NeurIPS 2024 · similarity 0.81)
- Reference-Based Post-OCR Processing with LLM for Precise Diacritic Text in Historical Document Recognition (AAAI 2025 · similarity 0.80)
- TAMER: Tree-Aware Transformer for Handwritten Mathematical Expression Recognition (AAAI 2025 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.