Persian Handwritten Digit Recognition
A CNN classifier for Persian (Farsi) handwritten digits — ۰۱۲۳۴۵۶۷۸۹ —
trained on 80,000 real handwritten digit images.
Accuracy: 99.8% on a held-out test set (200 images per class).
Model
- Architecture: Conv2D → Conv2D → MaxPool → Dense → Softmax (10 classes)
- Input: 28×28 grayscale image, normalized to [0, 1]
- Output: probability distribution over 10 digits (0–9)
- Framework: TensorFlow / Keras 3 (
digit_classifier_handwritten.keras) - Training data:
Mehdinmz/persian-handwritten-digits
Usage
from huggingface_hub import hf_hub_download
import tensorflow as tf
import numpy as np
path = hf_hub_download("Mehdinmz/persian-handwritten-digit-recognition", "digit_classifier_handwritten.keras")
model = tf.keras.models.load_model(path)
# img: 28x28 grayscale float32 in [0,1]
pred = model.predict(img.reshape(1, 28, 28, 1), verbose=0)
digit = int(np.argmax(pred)) # 0..9 (Persian digit: ۰۱۲۳۴۵۶۷۸۹)
Results
| Metric | Value |
|---|---|
| Test accuracy | 99.8% |
| Per-class (200 imgs each) | ≥ 99.5% for all classes |
Source
Full OCR pipeline (segmentation, CAPTCHA support, training code): github.com/mehdinmz/Persian-CAPTCHA-OCR
License
MIT © 2026 Mohammad Mehdi Namazian
- Downloads last month
- 10