🌫️ PyTorch Single Image Dehazing

Dark Channel Prior Implementation

A PyTorch implementation of CVPR 2009 "Single Image Haze Removal Using Dark Channel Prior"
✨ Supports batch-level operations for efficient processing ✨

Features • Installation • Quick Start • Parameters • Results • Citation

---

🎯 Features

• ⚡ Batch Processing: Efficient batch-level operations for multiple images
• 🔧 PyTorch Native: Fully implemented in PyTorch for GPU acceleration
• 📊 Intermediate Results: Access to dark channel, transmission map, and depth estimation
• 🎛️ Customizable: Adjustable parameters for different dehazing scenarios
• 🚀 Easy to Use: Simple API with minimal setup

📦 Installation

Dependencies

pip install torch torchvision numpy pillow

Required packages:

• pytorch - Deep learning framework
• numpy - Numerical computing
• PIL (Pillow) - Image processing

Clone Repository

git clone https://github.com/Magicboomliu/Pytorch-Version-of-Single-Image-Dehazing-using-Dark-Channel-Prior.git
cd Pytorch-Version-of-Single-Image-Dehazing-using-Dark-Channel-Prior

🚀 Quick Start

from dehaze_torch import DarkChannelPrior

# Initialize the dehazing model
dehaze_model = DarkChannelPrior(
    kernel_size=15,
    top_candidates_ratio=0.0001,
    omega=0.95,
    radius=40,
    eps=1e-3,
    open_threshold=True,
    depth_est=True
)

# Apply dehazing
dehaze_images, dc, airlight, raw_t, refined_transmission, depth = dehaze_model(image_data_tensor)

📝 Example Usage

import torch
from PIL import Image
from torchvision import transforms
from dehaze_torch import DarkChannelPrior

# Load and preprocess image
image = Image.open('fogg.png')
transform = transforms.Compose([
    transforms.ToTensor(),
])
image_tensor = transform(image).unsqueeze(0)  # Add batch dimension

# Initialize and run dehazing
dehaze_model = DarkChannelPrior()
dehazed_image, _, _, _, _, _ = dehaze_model(image_tensor)

# Save result
dehazed = transforms.ToPILImage()(dehazed_image.squeeze(0))
dehazed.save('output.png')

🎛️ Parameters

Parameter	Type	Default	Description
kernel_size	int	15	Size of the local patch for dark channel computation
top_candidates_ratio	float	0.0001	Ratio of top pixels to estimate atmospheric light
omega	float	0.95	Parameter to keep a small amount of haze (0-1)
radius	int	40	Radius for guided filter
eps	float	1e-3	Regularization parameter for guided filter
open_threshold	bool	True	Enable thresholding for transmission map
depth_est	bool	True	Enable depth estimation

🔍 Output Components

The model returns a tuple containing:

1. dehaze_images - Final dehazed images
2. dc - Dark channel of the input images
3. airlight - Estimated atmospheric light
4. raw_t - Raw transmission map
5. refined_transmission - Refined transmission map (after guided filtering)
6. depth - Estimated depth map

📊 Results

Visualization of Dehazing Process

The following image shows the complete dehazing pipeline including intermediate results:

From left to right: Original hazy image, dark channel, transmission map, refined transmission, and dehazed result

📚 References

This implementation is based on the following papers and repositories:

Papers:

• 📄 Single Image Haze Removal Using Dark Channel Prior - Kaiming He, Jian Sun, and Xiaoou Tang, CVPR 2009
• 📄 Guided Image Filtering - Kaiming He, Jian Sun, and Xiaoou Tang, ECCV 2010

Repositories:

• He-Zhang/image_dehaze
• anjali-chadha/dark-channel-prior-dehazing

📖 Citation

If you find this implementation helpful in your research, please consider citing:

@misc{zihualiu2023DCP,
  title={Pytorch-Version-of-Single-Image-Dehazing-using-Dark-Channel-Prior},
  author={Zihua Liu},
  howpublished={\url{https://github.com/Magicboomliu/Pytorch-Version-of-Single-Image-Dehazing-using-Dark-Channel-Prior}},
  year={2023}
}

📄 License

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