Researchers have developed DU2Net, a lightweight U-shaped network for underwater image enhancement that addresses color distortion, low contrast, and detail blur caused by light refraction and absorption. The method leverages a large-scale dataset of 11,739 real underwater images (DSUI) and integrates axial depthwise convolution and dense attention blocks to reduce computational complexity and parameter count, boosting processing speed and image quality.
Key takeaways
- DU2Net outperforms state-of-the-art methods like UDCP and CRUHL, improving UIQM by 0.367, UCIQE by 0.072, CCF by 26.165, and AG by 7.833.
- The network runs 8 times faster than UDCP, making it suitable for real-time applications in underwater robotics and inspection.
- A multi-color-space loss function combining RGB, LAB, and LCH spaces enhances color fidelity and contrast, aligning with human visual perception.
This approach offers a practical solution for autonomous underwater vehicles and remote operated vehicles, enabling clearer vision for navigation, target recognition, and ecological monitoring in challenging underwater environments.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-01-13 · “基于轻量级U形网络的颜色空间优化水下图像增强方法”