🔐 OCR ReCaptcha Solver (MB Bank CAPTCHA)

This project is an end-to-end OCR system designed to recognize CAPTCHA images—specifically from some Banks. It includes tools to scrape, train, and serve a model that can decode distorted characters from CAPTCHA images using deep learning.

---

📸 Overview

---

🧠 Features

• 🚀 CNN-based CAPTCHA recognition using PyTorch
• 🛠 Custom Dataset loader with multi-label classification
• 📥 CAPTCHA scraper from MB Bank login page
• 🌐 Flask API server for real-time predictions
• 📦 Clean and modular structure for training + inference

---

🗂️ Project Structure

ocr-recaptcha/
├── __pycache__/                # Python cache
├── mbbank_dataset/            # Folder for training/test CAPTCHA images
├── CONFIG.py                  # Configuration (image size, char set, model path)
├── MAIN.py                    # Main script to train or test the model
├── datasets.py                # Custom PyTorch Dataset class
├── mbbank_scrapping.py        # Script to scrape new CAPTCHA images
├── model.pth                  # Trained CNN model
├── requirements.txt           # Python dependencies
├── server.py                  # Flask server for API-based prediction
├── temp.png                   # Temporary file used for API testing

---

⚙️ Installation

git clone https://github.com/co-dev0909/ocr-recaptcha.git
cd ocr-recaptcha
pip install -r requirements.txt

---

📥 Scrape CAPTCHA Images

Run the scraping script to download new CAPTCHA images:

python mbbank_scrapping.py

Images are saved into mbbank_dataset/ (you can modify this inside the script).

---

🏋️ Train the Model

Ensure your image filenames represent the correct CAPTCHA labels (e.g., AB12cd.jpg):

python MAIN.py

This will train the CNN and save the weights to model.pth.

---

🔍 Test the Model

To test the model on all CAPTCHA images in a folder:

# Inside MAIN.py, you can call:
CrackLettesInt4().check_all('mbbank_dataset/test')

This will:

• Predict each CAPTCHA
• Print the result
• Move it to ./test/<prediction>.jpg

---

🌐 Run the Flask API

Start a REST API server for inference:

python server.py

🧪 Test API

Send a file:

curl -X POST -F "image=@mbbank_dataset/sample.jpg" http://localhost:8001/predict

Send base64:

POST /predict
{
  "image_base64": "data:image/png;base64,iVBORw0KGgoAAAANS..."
}

Response:

{
  "prediction": ["4Hd6Rf"]
}

---

🔠 Character Set

Supported characters (from CONFIG.py):

123456789ABDEFGHJMNRTYabdefghjmnqrty

• CAPTCHA Length: 6
• Image Size: 200x35

---

🧾 Requirements

torch
torchvision
flask
Pillow
numpy
shortuuid

Install via:

pip install -r requirements.txt

---

✅ Tips

• Customize CAPTCHA source URLs in mbbank_scrapping.py.
• Update image paths or save logic in MAIN.py and datasets.py if needed.
• To extend: Add beam search, augmentations, or use CRNN with CTC for harder CAPTCHAs.

---

🧑‍💻 Author

Maintained by co-dev0909.

---

📄 License

This project is open-source and available under the MIT License.