SAQR: A Paired Printed–Handwritten Arabic Line Dataset for Handwriting Recognition and Cross-Modal Retrieval

Paper: SAQR: A Paired Printed–Handwritten Arabic Line Dataset for Handwriting Recognition and Cross-Modal Retrieval
Journal: Scientific Data (Nature/Springer) — under review
Contact: mahmoud.salah@aun.edu.eg

---

Dataset Summary

SAQR is a large-scale paired Printed Ground Truth – Handwritten (GT-HW) Arabic dataset collected from 331 Arabic-speaking students aged 12–22. Each sample pairs a printed reference line image with the corresponding student handwriting, providing rich supervision for Arabic Handwritten Text Recognition (HTR) and related vision tasks.

Property	Value
Total pairs	2,263 GT-HW line pairs
Writers	331 students (age 12–22)
Gender	1,207 male / 1,056 female
Total words	25,352
Total characters	111,437 Arabic characters
Unique vocabulary	4,047 word types
Unique characters	47
Avg. line length	49.2 chars / 11.2 words
Splits	Train 1,575 / Val 335 / Test 353 (writer-independent)
Language	Arabic (Modern Standard Arabic / formal prose)
License	CC-BY 4.0

Dataset Preview

Sample GT–HW line pairs from SAQR, showing the range of handwriting styles across the writer population (printed ground truth on top, student handwriting below):

Supported Tasks

Task	Description
Handwritten Text Recognition (HTR)	Transcribe hw_image → Arabic text
Gender Classification	Predict writer gender from hw_image
GT-HW Cross-Modal Matching	Match gt_image ↔ hw_image across 353 candidates

Dataset Structure

Each sample contains:

• pair_id — unique integer identifier
• gt_image — image of the printed Arabic text line (Ground Truth)
• hw_image — image of the handwritten copy (student writing)
• text — Arabic transcription string
• category — writer gender (0=boys / 1=girls)
• split — train / val / test
• student_id — anonymized form identifier
• form_creator — form template identifier

Loading the Dataset

Full dataset

from datasets import load_dataset

ds = load_dataset("mahmoudsalah01/SAQR")
train = ds["train"]
val   = ds["validation"]
test  = ds["test"]

# Access a sample
sample = train[0]
print(sample["text"])        # Arabic transcription
sample["gt_image"].show()    # Printed GT line
sample["hw_image"].show()    # Handwritten line

HTR task (HW image → text)

from datasets import load_dataset
ds = load_dataset("mahmoudsalah01/SAQR", split="train")
for sample in ds:
    image = sample["hw_image"]   # Input image
    label = sample["text"]       # Target transcription

Gender classification task

from datasets import load_dataset
ds = load_dataset("mahmoudsalah01/SAQR", split="train")
for sample in ds:
    image  = sample["hw_image"]    # Input: handwriting image
    gender = sample["category"]    # 0=boys, 1=girls

Benchmark Results

Task 1 — Handwritten Text Recognition

Method	CER ↓	WER ↓
Tesseract 4.0 (zero-shot)	0.565	1.061
EasyOCR (zero-shot)	0.485	1.045
TrOCR-Base (fine-tuned, ours)	0.533	0.721

Task 2 — Gender Classification from Handwriting

Method	Accuracy	F1-Macro
ViT-Base (fine-tuned, ours)	73.5%	0.71

Task 3 — GT-HW Cross-Modal Matching (pool = 353)

Method	R@1	R@10	MRR
CLIP ViT-B/32 (zero-shot)	0.6%	7.1%	3.2%
DINOv2 Siamese	1.7%	11.0%	5.6%
Fine-tuned CLIP	5.9%	26.3%	12.9%
Siamese ViT (InfoNCE)	5.4%	36.5%	15.0%
Siamese ViT + Hard Neg (ours)	10.8%	31.7%	17.7%

Collection Methodology

Students copied text lines from printed forms (Arabic formal prose, sourced from newspaper articles) onto A4 paper sheets. Forms were scanned at 300 DPI, line-segmented, and paired by form index. Transcriptions were extracted from original .docx form templates. Splits are writer-independent: no writer appears in more than one split.

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Contact

Mahmoud Salah — mahmoud.salah@aun.edu.eg
For questions or issues, please open a Discussion.