https://github.com/xingyumex/UNIR-Net
Enhancing underwater images with non-uniform illumination (NUI) is crucial for improving visibility and visual quality in marine environments, where image degradation is caused by significant absorption and scattering effects. However, traditional model-based methods are often ineffective at capturing the complex illumination variations present in such images, resulting in limited visual improvements. On the other hand, learning-based approaches have shown promising results but face challenges due to the lack of specific datasets designed to effectively address the non-uniform illumination problem. To overcome these challenges, the Underwater Non-uniform Illumination Restoration Network (UNIR-Net) is introduced, a novel method that integrates illumination enhancement and attention blocks, along with visual refinement and contrast correction modules. This approach is specifically designed to mitigate the scattering and absorption effects that cause light attenuation in underwater environments. Additionally, the Paired Underwater Non-uniform Illumination (PUNI) dataset is presented, a paired resource that facilitates the restoration of underwater images under non-uniform illumination conditions. Extensive experiments conducted on the PUNI dataset and the large-scale real-world Non-Uniform Illumination Dataset (NUID), which contains underwater images with non-uniform illumination, demonstrate the robust generalization ability of UNIR-Net. This method outperforms existing approaches in both quantitative metrics and qualitative evaluations. Furthermore, UNIR-Net not only significantly enhances the visual quality of images but also improves performance in advanced computer vision tasks, such as semantic segmentation in underwater environments, highlighting its broad applicability and potential impact. The code of this method is available at