🔓 CAPTCHA Breaker

Breaking CAPTCHAs with 98%+ accuracy using Convolutional Recurrent Neural Networks and CTC loss

---

🌟 Headlines

• 🎯 98%+ character accuracy on synthetic CAPTCHAs
• ⚡ GPU-accelerated training with mixed precision
• 🔄 End-to-end pipeline from data generation to deployment
• 🧠 Full CTC evaluation
• 🎨 Synthetic data generation - no manual labeling required

🚀 Quick Start

Get up and running in minutes:

1️⃣ Install Dependencies

pip install tensorflow tensorflow-addons omegaconf hydra-core tqdm matplotlib captcha pillow

2️⃣ Generate Training Data

python generate_synthetic_dataset.py -n 20000

Creates 20,000 synthetic CAPTCHA images like ABC123_001.png

3️⃣ Train the Model

python train.py

Auto-detects GPU based on TensorFlow, trains with mixed precision, saves best model

4️⃣ Test Your Model

python evaluate.py
python predict.py path/to/your/captcha.png

---

📖 How It Works

🧠 The Architecture

Our CRNN (Convolutional Recurrent Neural Network) combines three powerful components:

📸 CAPTCHA Image → 🔍 CNN Feature Extractor → 🔄 Bidirectional LSTM → 📝 CTC Decoder → ✨ Text Output

1. CNN Backbone 🔍

• ResNet-inspired feature extractor
• Batch normalization + ReLU activation
• Progressive max pooling for spatial reduction
• Converts images to rich feature representations

2. Sequence Modeling 🔄

• Bidirectional LSTM layers capture left-to-right AND right-to-left context
• Handles variable-length sequences automatically
• Dropout prevents overfitting

3. CTC Magic ✨

• Connectionist Temporal Classification eliminates need for character-level alignment
• Handles variable-length outputs elegantly
• Proper decoding removes duplicates and blank tokens

🎯 Why CTC is Crucial

❌ Traditional Approach:

Requires: [A][B][C][1][2][3] ← Exact alignment needed

✅ CTC Approach:

Handles: [A][A][_][B][C][_][1][2][3][3][_] ← Automatic alignment
         ↓ CTC Decoding ↓
Output:  ABC123

---

🎯 Results & Performance

📊 Accuracy Metrics

Metric	Score	Description
Character Accuracy	99.3%	Individual character recognition
Sequence Accuracy	97.9%	Complete CAPTCHA solved correctly
Training Time	<1 hour	On RTX 3050 (mixed precision)
Inference Speed	~10ms	Per image on GPU

🏆 Benchmarks

Training Performance:

• Dataset: 20,000 synthetic CAPTCHAs
• Convergence: 20-40 epochs (early stopping)
• Memory Usage: ~2GB GPU memory
• Speed: 40% faster with mixed precision

📈 Learning Curves

The model typically shows:

• Rapid initial learning (epochs 1-10)
• Gradual improvement (epochs 10-30)
• Convergence with early stopping

---

🐛 Troubleshooting

Common Issues:

1. CUDA out of memory:

   # Reduce batch size in config.yaml
   batch_size: 64  # or 32

2. Mixed precision errors:

   # Disable for older GPUs
   mixed_precision: false

---

📜 License

This project is licensed under the MIT License - see LICENSE.txt for details.

TL;DR: Use it freely for educational and commercial purposes! 🎉