LVRNet: Lightweight Image Restoration for Aerial Images under Low Visibility

paper | dataset | demo | slides | poster | code | webpage

AAAI 2023 Student Abstract and Poster Program Submission

LVRNet, short for Low-Visibility Restoration Network, is a method that can effectively recover high-quality images from degraded images taken in poor visual conditions. Although we have tested our work for two degrading factors combined: low-light and haze, you can use this codebase and run experiments for other degrading factors as well using the instructions given below.

Method Overview

Quick Start

1. Install Environment

git clone https://github.com/Achleshwar/lvrnet.git
cd lvrnet 
pip install -r requirements.txt

2. Download Dataset

We have used public dataset AFO and generated our dataset - Low-Vis AFO, by adding low visibility conditions. You can download it here.

3. Demo using pretrained Weights

For a quick demo, you can use our pretrained weights and run them on a demo images using src/lvrnet-notebook.ipynb.

Download the pretrained weights from here and change model_wts path in the notebook.

4. Reproducing the results

## train from scratch
python train.py --epochs 50 --data_dir <path to dataset> --log_dir <path to save weights> --perloss --edgeloss --fftloss

Citation

If you find this work useful, please cite our paper:

@misc{pahwa2023lvrnet,
      title={LVRNet: Lightweight Image Restoration for Aerial Images under Low Visibility}, 
      author={Esha Pahwa and Achleshwar Luthra and Pratik Narang},
      year={2023},
      eprint={2301.05434},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Acknowledgement

We would like to thank the authors of FFANet, NAFNet and MC-Blur for their codebase. We have used their codebase as a starting point for our work.

TODOs

• Add results on OOD images
• Add link to dataset
• Add link to project page