Image Enhancement: A Necessity for Effective Underwater Object Detection?

TMLR Paper6130 Authors

Abstract:

Underwater vision is essential for applications such as marine engineering, aquatic robotics,
and environmental monitoring. However, severe image degradation caused by light absorption and scattering often compromises object detection (OD) performance. Although underwater image enhancement (UIE) intuitively seems beneficial for restoring visual information and improving detection accuracy, its actual impact remains unclear. This work systematically evaluates state-of-the-art enhancement models and investigates their effects on underwater OD to answer the key question: "Is UIE necessary for accurate OD?" We conducted a systematic evaluation of 20 representative UIE algorithms—spanning traditional methods, convolutional neural networks (CNNs), generative adversarial networks (GANs), Transformers, and Diffusion models. These methods are applied to two benchmark datasets, RUOD and URPC2020, producing 21 domain variants per dataset (raw +
20 enhanced). To rigorously assess the effect of enhancement on detection, we trained five object detectors on each domain, resulting in 210 unique model configurations (5 detectors × 21 domains × 2 datasets). Our findings reveal that, contrary to intuitive expectations, most enhancement techniques actually degrade detection accuracy. Only well-designed methods, such as diffusion-based approaches that preserve key low-level features without introducing artificial distortions, can minimize this negative impact. These results provide critical insights into the role of enhancement in underwater vision and highlight important considerations for future research.

Submission Length: Regular submission (no more than 12 pages of main content)

Previous TMLR Submission Url: https://openreview.net/forum?id=f6xl3Q442O

Changes Since Last Submission: The font issue in the initial version has been corrected in this submission. In addition, the figure captions and related works section have been simplified to ensure the paper fits within the 12-page limit. No other technical content has been altered.

Assigned Action Editor: ~Jianbo_Jiao2

Submission Number: 6130