Global Context Meets Local Quality: A Mamba-Enhanced Detector for Underwater Imagery

Abstract

Underwater object detection faces severe challenges caused by light attenuation, scattering, spatially varying turbidity, and boundary blur, which weaken object-related visual signals and reduce localization reliability. This letter presents MED, a Mamba-Enhanced Detector for degradation-aware underwater object detection, achieving 82.0% average mAP@0.5 and 48.7% average mAP@0.5:0.95 across three underwater datasets, with 15.5 M parameters and 30.1 GFLOPs. Three problem-driven modules address specific degradation factors: a Mamba-enhanced backbone for selective global-context filtering under attenuated visibility, Dynamic Gated Attention Fusion (DGAF) for adaptive multi-scale feature fusion under non-uniform turbidity, and a Scattering-Aware Quality Estimation (SAQE) head for localization-quality calibration under scattering-induced boundary uncertainty. Extensive experiments on RUOD, UDD, and URPC2020 demonstrate that MED consistently improves detection accuracy over YOLOv11 s and recent underwater detectors, providing a practical accuracy-complexity trade-off for robust underwater perception.