Arabic handwritten digits, read accurately.
Raqam OCR turns handwritten Arabic numerals on forms, cheques and invoices into clean digital data, using convolutional neural networks trained on 70,000 real handwriting samples.
The problem
Handwritten numbers still slow down paperwork.
Banks, schools, clinics and public offices still process forms and cheques filled in by hand. Arabic digits vary widely between writers, and manual re-typing is slow and error-prone, where one wrong digit can change an amount or an ID.
handwritten Arabic digit images used to train and test our models.
different writers contributed samples, capturing real variation in handwriting.
test accuracy of our best model: a CNN with data augmentation and dropout.
Source: the founding research paper (IEEE, 2025), using the Arabic Handwritten Digits Database (AHDD1).
How it works
From scanned page to verified numbers.
Raqam OCR slots into existing scanning workflows and returns structured, validated values.
Scan
Forms or cheques are scanned or photographed with any standard device.
Segment
Number fields are located and split into individual digits.
Recognise
A CNN trained with augmentation and dropout classifies each digit 0–9.
Verify
Low-confidence digits are flagged for a quick human check before export.
Product
Accuracy you can build a workflow on.
Tested on 10,000 unseen digits, the model stays robust across very different handwriting styles.
CNN with augmentation
The best-performing architecture in our study, robust to slant, thickness and style.
Lightweight options
Simpler MLP and fully connected networks also performed well at lower computational cost, for edge devices.
Confidence flags
Every digit carries a confidence score so uncertain values are reviewed, not guessed.
Easy integration
Outputs clean structured data ready for banking, records or ERP systems.
Who it's for
For anyone who still receives numbers by hand.
Banks
Read handwritten cheque amounts and account numbers faster.
Public services & schools
Digitise application forms, exam sheets and registers.
Healthcare & logistics
Capture handwritten IDs, quantities and readings without re-typing.
The research
Benchmarked across multiple deep learning models.
- Uses the AHDD1 dataset: 70,000 images, 60,000 for training and 10,000 for testing.
- Digits 0–9 written by 700 contributors from diverse backgrounds.
- CNN with data augmentation and dropout achieved the highest accuracy: 99.32%.
- MLP and fully connected networks performed well at lower computational cost; more complex models like ResNet gave only slight differences.
Zraqou, J. S., Zeer, M., & Sabateen, M. (2025). Advanced Deep Learning Techniques for Arabic Handwritten Digit Classification. IEEE ICCIAA 2025.
Team
Built by computer engineers.
Founded in 2025 by a student team from the faculty below.
Faculty of Engineering & Information Technology
Deep learning research, model training and product engineering.
FAQ
Questions we hear often.
Does it read Eastern Arabic numerals (٠–٩)?
Yes. The models are trained specifically on handwritten Arabic digits 0–9.
How accurate is it?
Our best model achieved 99.32% accuracy on 10,000 test images. Low-confidence digits are flagged for review.
Can it run offline?
Lighter MLP-based models are suitable for local or edge deployment where connectivity is limited.
Can we pilot it on our forms?
Yes. Share sample documents and we will run a pilot on your real workflow.
Stop re-typing handwritten numbers.
We are onboarding pilot partners in banking, education and public services.