PersianDigitOCR

%%writefile README.md

PersianDigitOCR

Overview

PersianDigitOCR is a deep learning model designed for recognizing handwritten Persian (Farsi) digits (۰-۹) from images. Built by fine-tuning the MobileNetV2 architecture pre-trained on ImageNet, this model achieves a validation accuracy of 99.75% on a subset of the HODA dataset. It is optimized for 96x96 pixel grayscale images and outputs Persian digit characters using Unicode mappings (U+06F0 to U+06F9).

• Pipeline Tag: image-classification
• Date: October 08, 2025

Model Details

• Base Model: MobileNetV2 (from google/mobilenet_v2_1.0_224)
• Input: Grayscale images resized to 96x96 pixels
• Output: Predicted digit as a Persian Unicode character (e.g., '۰' for 0)
• Training Data: Subset of 20,000 images from the HODA dataset (80% training, 20% validation)

Usage

import requests
# Step 1: Download the model from Hugging Face
model_filename = 'PersianDigitOCR.keras'  
repo_id = 'shaghxyegh/PersianDigitOCR'
file_url = f"https://huggingface.co/{repo_id}/resolve/main/{model_filename}"

print(f"Downloading model from {file_url}...")
response = requests.get(file_url)
if response.status_code == 200:
    with open(model_filename, 'wb') as f:
        f.write(response.content)
    print(f"Model downloaded successfully as '{model_filename}'")
else:
    raise ValueError(f"Failed to download model. Status code: {response.status_code}. Check if the file '{model_filename}' exists in your repo at {file_url}")

# Step 2: Load the model
model = load_model(model_filename)
print("Model loaded successfully!")

Performance

Validation Accuracy: 99.75% on 4,000 validation images Training Accuracy: 99.69% (Epoch 30) Context: Trained with batch size 32, 30 epochs, Adam optimizer, and data augmentation (rotation, zoom, shift).