Image enhancement: A necessity for effective underwater object detection?

Underwater vision plays a critical role in engineering applications such as marine inspection, aquatic robotics, and environmental monitoring, where reliable object detection is essential. However, severe image degradation caused by light absorption and scattering often limits detection performance. Although underwater image enhancement (UIE) is widely used to improve visual quality, its actual impact on downstream detection tasks remains unclear. This work presents a large-scale empirical analysis investigating how enhancement-induced distribution shifts influence deep learning-based object detection models. We systematically evaluate 20 representative UIE methods, including traditional approaches, convolutional neural networks, generative adversarial networks, Transformers, and diffusion models, across two benchmark underwater datasets, the Real-world Underwater Object Detection dataset (RUOD) and the Underwater Robot Picking Contest 2020 dataset (URPC2020). To quantify engineering impact, five object detectors are trained and evaluated under 210 experimental configurations (5 detectors ×21 domains ×2 datasets), enabling comprehensive analysis of performance trends across object scale, turbidity levels, and texture complexity. Our results suggest that, contrary to common assumptions, many enhancement methods are associated with reduced detection accuracy. Further analysis indicates that enhancement-induced color shifts, structural distortions, and feature-distribution changes may contribute to performance degradation. Distribution-shift analysis reveals strong correlations between pixel-, texture-, and perceptual-level changes and detection outcomes across detectors and datasets. Among the evaluated paradigms, diffusion-based enhancement exhibits the most stable behavior and remains more consistent with the original RAW image distribution than other enhanced methods.