UNIR-Net: A Novel Approach for Restoring Underwater Images with Non-Uniform Illumination Using Synthetic Data

TL;DR

UNIR-Net is a new deep learning model that effectively restores underwater images with complex non-uniform illumination, using a novel synthetic dataset for training and achieving superior results in image quality and downstream tasks.

Contribution

The paper introduces UNIR-Net, a novel network architecture for underwater image restoration under non-uniform illumination, and the PUNI dataset for targeted training and evaluation.

Findings

UNIR-Net outperforms existing methods in quantitative metrics.

UNIR-Net improves visual quality of underwater images.

Enhances downstream tasks like underwater semantic segmentation.

Abstract

Restoring underwater images affected by non-uniform illumination (NUI) is essential to improve visual quality and usability in marine applications. Conventional methods often fall short in handling complex illumination patterns, while learning-based approaches face challenges due to the lack of targeted datasets. To address these limitations, the Underwater Non-uniform Illumination Restoration Network (UNIR-Net) is proposed. UNIR-Net integrates multiple components, including illumination enhancement, attention mechanisms, visual refinement, and contrast correction, to effectively restore underwater images affected by NUI. In addition, the Paired Underwater Non-uniform Illumination (PUNI) dataset is introduced, specifically designed for training and evaluating models under NUI conditions. Experimental results on PUNI and the large-scale real-world Non-Uniform Illumination Dataset (NUID) show that UNIR-Net achieves superior performance in both quantitative metrics and visual outcomes. UNIR-Net also improves downstream tasks such as underwater semantic segmentation, highlighting its practical relevance. The code of this method is available at https://github.com/xingyumex/UNIR-Net