🎨 Image Colorization using Deep Learning

A dual-model image colorization system that converts black & white images into realistic color images using deep learning.

This project demonstrates both:

• 🧠 A custom-trained model (built from scratch)
• 🎯 A state-of-the-art pretrained model (Zhang et al., ECCV 2016)

---

🚀 Live Demo

👉 Deployable via Streamlit Cloud (First run may take a few seconds to download the pretrained model)

Click for Live Demo

---

✨ Features

• 🔄 Dual Model System

  • Local trained model (fast, lightweight)
  • Pretrained model (high-quality results)

• 🆚 Comparison Mode

  • View Original vs Local vs Pretrained side-by-side

• 🎨 Color Boosting

  • Enhances saturation for better visual output in local model

• 📥 Smart Model Handling

  • Pretrained model is downloaded automatically at runtime
  • Avoids GitHub file size limits

• 📱 Mobile Friendly UI

  • Optimized layout for all screen sizes

---

🧠 Models Used

🔹 Local Model (Custom Trained)

• Trained on Oxford-IIIT dataset
• Input: Grayscale (L channel)
• Output: Color channels (AB)
• Architecture: CNN
• Loss: MAE
• Activation: Tanh (full color range)

«Built to demonstrate deep learning pipeline under limited hardware constraints»

---

🔹 Pretrained Model

Based on:

"Colorful Image Colorization" — Zhang et al. (ECCV 2016)

• Trained on large-scale datasets
• Implemented using OpenCV DNN (Caffe)
• Produces high-quality realistic results

---

⚙️ Pretrained Model Handling

Due to GitHub’s 100MB file size limit, the pretrained model is not stored in the repository.

Instead:

• 📥 It is automatically downloaded at runtime
• ⚡ Cached after first use
• 🚀 Ensures smooth deployment on Streamlit Cloud

---

🧪 How It Works

1. Convert image → LAB color space
2. Extract L channel (grayscale)
3. Predict AB channels using model
4. Merge L + AB
5. Convert back to RGB

---

🖼️ Visuals

🔹 Pipeline

🔹 LAB Color Space

🔹 Network Architecture

---

📁 Project Structure

Image-Colorization/
│
├── app.py
├── colorize.py
├── requirements.txt
│
├── models/
│ ├── local-trained/
│ │ └── colorization_model.keras
│ │
│ ├── pre-trained/ # auto-downloaded at runtime
│ │
│ └── model_training/
│ ├── core_model.ipynb
│ ├── colorization_training.ipynb
│ └── colorization_example.ipynb
│
├── assets/
└── README.md

---

▶️ Run Locally

git clone https://github.com/your-username/Image-Colorization.git

cd Image-Colorization

python -m venv venv
source venv/bin/activate

pip install -r requirements.txt

streamlit run app.py

---

⚠️ Limitations

• Local model is less accurate than pretrained
• Colorization is inherently ambiguous
• Results depend on image context

---

📚 References

• Zhang, R., Isola, P., & Efros, A. A. (2016). Colorful Image Colorization

• OpenCV DNN Module

• TensorFlow / Keras

---