LD-Net: Lightweight Dehazing Network Using Convolutional Autoencoder for Image Dehazing

LD-Net is a deep learning-based algorithm designed to enhance visibility in hazy images and videos, improving visual quality and clarity. This project implements a lightweight dehazing network based on a convolutional autoencoder (CAE) architecture.

Key Features

• Achieved a 40% improvement in structural similarity index (SSIM) compared to baselines
• Reduced model size by 70% while maintaining performance
• Integrated PyTorch model with a React.js frontend via a Flask API
• Enables real-time image processing with under 2-second latency

Project Structure

The project consists of two main components:

1. Frontend (React.js)
2. Backend (Python Flask API)

Installation and Setup

Frontend

cd frontend
npm install
npm run dev

Backend

cd Backend
pip install -r requirements.txt
python main.py

Virtual Environment Setup

For Linux-based systems:

python -m venv .venv
source .venv/bin/activate

For Windows-based systems:

python -m venv .venv
source .venv/Scripts/activate

Required Python Modules

Install the following modules in your virtual environment:

pip install torch opencv-python numpy tqdm matplotlib

Note: Make sure to compile and run the project using the virtual environment kernel.

Technical Details

LD-Net utilizes the encoder-decoder structure of a Convolutional Autoencoder (CAE) to learn a latent representation that captures the haze-free content of the image. The decoder then reconstructs the dehazed image from this latent representation.

Key advantages of LD-Net include:

• Lightweight architecture suitable for real-time applications
• Computational efficiency
• Improved image quality metrics (PSNR and SSIM) compared to hazy input images

Evaluation

The network has been evaluated on a benchmark hazy image dataset, demonstrating its effectiveness in improving image quality metrics like PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index).

Future Work

• Further optimization of the model for even faster processing
• Expansion of the dataset to improve generalization
• Integration with mobile platforms for on-device dehazing

Contributors

Ashutosh Jha Ayush Sehgal