Re-Degradation and Contrastive Learning for Zero-shot Underwater Image Restoration
Nisha Varghese (IIT Madras),* Rajagopalan N Ambasamudram (Indian Institute of Technology Madras)
The 34th British Machine Vision Conference
Abstract
Restoring underwater (UW) images is an important task in ocean exploration applications and is quite challenging due to its fundamental ill-posedness. Traditional methods for UW image restoration struggle when there is a mismatch between the adopted prior and actual scene conditions. Deep models require large-scale paired or unpaired real-world data for training which are scarce in the UW scenario; synthetic datasets typically suffer from domain-shift issues. To alleviate the limitations of prior-based and data-driven UW restoration methods, "zero-shot" approach is an attractive solution. In this paper, we propose an UnderWater Zero-shot image Restoration method (UWZR) by harnessing the physical model for UW image formation. A re-degradation strategy is introduced to generate another UW image that respects the same image formation model. The network is optimized to disentangle the input UW image in such a manner that the relationships between the components of the input UW image and the re-degraded image are satisfied. A contrastive learning strategy is added that ensures that the restored image is pulled closer to a clean image and pushed far away from the UW image in the representation space. Extensive experiments on four real UW datasets establish the superiority of our proposed UWZR over prior art for UW image restoration.Video
Citation
@inproceedings{Varghese_2023_BMVC,
author = {Nisha Varghese and Rajagopalan N Ambasamudram},
title = {Re-Degradation and Contrastive Learning for Zero-shot Underwater Image Restoration},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0544.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by