🎨 Image Colorization Tool using OpenCV & Deep Learning

Transform old black-and-white photos into vibrant, realistic color images using Deep Learning + OpenCV, wrapped in a modern Tkinter GUI with advanced post-processing enhancements. This project leverages a pretrained CNN-based colorization model and applies multiple post-processing techniques to significantly improve visual quality.

---

✨ Features

• 🖼️ Black & White Image Colorization
• 🧠 Deep Learning-based Color Prediction (OpenCV DNN)
• 🎛️ Advanced Post-Processing Pipeline
• 🧴 Noise reduction & edge-preserving smoothing
• 🌈 Contrast enhancement using CLAHE
• 🎨 Controlled saturation boosting
• 🧑 Skin tone detection & correction
• 🔍 Advanced sharpening for fine details
• 📈 Super-resolution upscaling
• 🧵 Multi-threaded processing (UI never freezes)
• 💾 Save colorized images in high quality
• 🎨 Modern dark-themed GUI built with Tkinter

---

🧠 Methodology

1. Input image is converted from BGR → LAB color space
2. L channel (grayscale) is fed to a pretrained CNN
3. Model predicts A & B color channels
4. LAB image is reconstructed and converted back to BGR
5. Post-processing improves realism and perceptual quality

---

🛠️ Tech Stack

• Python
• OpenCV (DNN module)
• NumPy
• Tkinter
• Pillow (PIL)

---

📂 Project Structure

Image-Colorization/
│
├── Models/
│ ├── colorization_deploy_v2.prototxt
│ ├── colorization_release_v2.caffemodel
│ └── pts_in_hull.npy
│
├── colorization_gui.py
├── requirements.txt
└── README.md

---

🚀 Setup

1️⃣ Clone the repository

git clone "url"
cd image-colorization-tool

2️⃣ Install dependencies

pip install -r requirements.txt

3️⃣ Download the pretrained model

Place the following files inside the Models/ directory:

• colorization_deploy_v2.prototxt

• colorization_release_v2.caffemodel

• pts_in_hull.npy

• ℹ️ See the References section below for official download links.

---

▶️ Usage

Run the application:

python colorization_gui.py

Workflow

1. 📁 Click Browse Image
2. ✨ Click Colorize
3. 🎉 Preview the colorized output
4. 💾 Save the result

---

🎯 Post-Processing Pipeline

To improve realism, the following enhancements are applied after colorization:

1. 🔇 Noise reduction (Non-local Means)
2. 🌗 Contrast enhancement (CLAHE on L channel)
3. 🌈 Saturation boost (HSV space)
4. 🧠 Edge-preserving smoothing (Bilateral Filter)
5. 🧑 Skin tone detection & correction
6. 🔍 Advanced sharpening
7. 📈 Super-resolution upscaling

This significantly reduces:

• Washed-out colors
• Color bleeding
• Flat contrast
• Unrealistic skin tones

---

🖼️ Preview (Results)

Original vs Colorized Output

Original Black & White Image	Colorized Image

---

📊 Model Evaluation Metrics

The performance of the image colorization model was evaluated using standard image quality and perceptual similarity metrics. The following are the average results obtained across the evaluation dataset:

Metric	Value	Description
PSNR (Peak Signal-to-Noise Ratio)	19.86 dB	Measures reconstruction quality; higher values indicate better fidelity
SSIM (Structural Similarity Index)	0.9086	Evaluates perceptual and structural similarity (closer to 1 is better)
MSE (Mean Squared Error)	933.44	Measures average pixel-wise error; lower is better
Color Correlation	0.9610	Indicates how closely the predicted colors match the ground truth

📈 Metrics Visualization

The following visualization provides a graphical overview of the evaluation metrics, helping to better understand the model’s performance across different quality measures:

📌 Interpretation

• High SSIM and Color Correlation values indicate strong preservation of image structure and realistic color distribution.
• The PSNR value aligns with typical learning-based colorization models, which favor perceptual realism over pixel-level accuracy.
• Overall, the metrics confirm that the model produces visually coherent and perceptually convincing colorized images.

---

🧪 Known Limitations

• Colors are predicted, not restored — results may vary
• Rare objects may receive inaccurate colors
• Works best with clear grayscale or B&W images

---

📚 References

• Zhang, R., Isola, P., & Efros, A. A.
  Colorful Image Colorization
  https://arxiv.org/abs/1603.08511

• Official Pretrained Model Repository
  GitHub - richzhang/colorization

• OpenCV Deep Neural Network (DNN) Module
  OpenCV DNN Docs

• OpenCV Image Processing Documentation
  OpenCV Docs

• Non-Local Means Denoising (OpenCV)
  Tutorial: Non-Local Means Denoising

• CLAHE – Contrast Limited Adaptive Histogram Equalization
  Tutorial: CLAHE

• Model Files for Colorization (Caffe-based)

  • colorization_deploy_v2.prototxt (network architecture)
    Download Link
  • colorization_release_v2.caffemodel (pre-trained weights)
    Download Link
  • pts_in_hull.npy (cluster centers for ab channels)
    Download Link

👨‍💻 Developed By - @Arijit2175