<p align="center"> <h1 align="center">Phaseformer : Phase-based Attention Mechanism for Underwater Image Restoration and Beyond</h1> <p align="center"> <a href="https://mdraqibkhan.github.io">Md Raqib Khan</a> · <a href="https://scholar.google.com/citations?user=UcUMYe8AAAAJ&hl=en&oi=sra">Anshul Negi</a> · <a href="https://scholar.google.com/citations?user=WwdYdlUAAAAJ&hl=en&oi=sra">Asuhtosh Kulkarni </a> · <a href="https://scholar.google.com/citations?user=HgX8wb8AAAAJ&hl=en&oi=sra">Shruti S. Phutke</a> · <a href="https://visionintelligence.github.io">Santosh Kumar Vipparthi</a> · <a href="https://www.scss.tcd.ie/~muralas/">Subrahmanyam Murala</a> </p> <h3 align="center">WACV-2025</h3> <h3 align="center"><a href="https://arxiv.org/pdf/2412.01456">Paper</h3> <div align="center"></div> </p> <div align="center"> <img src="last_model.png" alt="Phaseformer : Phase-based Attention Mechanism for Underwater Image Restoration and Beyond" width="100%"> <h1>Phaseformer : Phase-based Attention Mechanism for Underwater Image Restoration and Beyond</h1> </div> <div align="center"></div> </p> <div align="center"> <img src="tsne_visuals.png" alt="t-SNE Visualization" width="100%"> <h1>t-SNE Visualization of the Amplitude and Phase of Clean and Degraded Images</h1> <p> The separate clusters for clean and degraded amplitude show that there is more effect of degradation on amplitude content compared to phase content, which has overlapping clusters for clean and noisy images. </p> </div>

Evaluation

To evaluate the model on different datasets using the provided checkpoints and sample degraded images.

Dataset and Checkpoint Structure

• Sample degraded images for testing: Available in dataset/dataset_name/.
• Checkpoints for evaluation: Provided in checkpoints/dataset-name/.
• Results storage: After successful execution, the results will be saved in the results/dataset-name/ folder.

Folder Overview

├── dataset
│   ├── UIEB
│   ├── U-45
│   ├── SQUID
│   ├── UCCS
│   ├── UFO-120
│   ├── Low_light
├── checkpoints
│   ├── UIEB
│   ├── UFO-120
│   ├── Low_light
├── results
│   ├── UIEB
│   ├── U-45
│   ├── SQUID
│   ├── UCCS
│   ├── UFO-120
│   ├── Low_light

Running the Evaluation

To evaluate the model on different datasets, follow the instructions below for each specific dataset:

UIEB Dataset Evaluation

Run the following command to evaluate the model on the UIEB dataset:

python test.py --dataset datasets/UIEB/  --checkpoints_path /checkpoints/UIEB/ --save_path Results/UIEB 

U-45 Dataset Evaluation

Run the following command to evaluate the model on the U-45 dataset:

python test.py --dataset dataset/U-45/ --checkpoints_path /checkpoints/UIEB/ --save_path Results/U-45

SQUID Dataset Evaluation

Run the following command to evaluate the model on the SQUID dataset:

python test.py --dataset dataset/SQUID/ --checkpoints_path /checkpoints/UIEB/ --save_path Results/SQUID

UCCS Dataset Evaluation

Run the following command to evaluate the model on the UCCS dataset:

python test.py --dataset dataset/UCCS/ --checkpoints_path /checkpoints/UIEB/ --save_path Results/UCCS

UFO-120 Dataset Evaluation

Run the following command to evaluate the model on the UCCS dataset:

python test.py --dataset dataset/UFO-120/ --checkpoints_path /checkpoints/UFO-120/ --save_path Results/UFO-120

Low-light Dataset Evaluation

Run the following command to evaluate the model on the UCCS dataset:

python test.py --dataset dataset/UCCS/ --checkpoints_path /checkpoints/Low_light/ --save_path Results/Low_light

Traing

1. Structure of data for training should be like

uw_data/   # here uw_data can be any underwater datsets folder like UIEB,UFO-120 etc.
   ├── train/
   │   ├── a/  # Input images
   │   └── b/  # Reference (ground truth) images
   └── test/
       ├── a/  # Input images
       └── b/  # Reference (ground truth) images

3. run

  pyhthon train.py

Citation

If you find this work helpful, please reference it as follows:

 @article{khan2024phaseformer,
  title={Phaseformer: Phase-based Attention Mechanism for Underwater Image Restoration and Beyond},
  author={Khan, MD and Negi, Anshul and Kulkarni, Ashutosh and Phutke, Shruti S and Vipparthi, Santosh Kumar and Murala, Subrahmanyam},
  journal={arXiv preprint arXiv:2412.01456},
  year={2024}
}

Acknowledgements

Special thanks to the awesome repositories UIPTA and Spectroformer, which made this project possible.