UEAOD: Underwater Image Enhancement-Assisted Object Detection Guided by Frequency Decomposition and Physical Consistency

Underwater vision is central to marine robotics and ecological monitoring, yet turbidity, color cast, and backscatter degrade appearance and confound object boundaries. Existing enhancement–detection pipelines often optimize perceptual quality, keep the enhancement branch active during inference, and process features uniformly in the spatial domain, yielding residual artifacts, domain sensitivity, and nontrivial latency. We propose underwater enhancement-assisted object detection (UEAOD), a lightweight detector that unifies frequency-aware factorization with physics-guided self-supervision. Specifically, a Fourier frequency decoupling module adaptively splits each feature map into low- and high-frequency (HF) components using a learnable threshold optimized with a straight-through estimator (STE). Building on this base–detail separation, a contextual dual-stream network (CDSN) compatible with the path aggregation network (PAN) refines the low-frequency (LF) base through a top-down path and consolidates HF detail with content-aware upsampling. Then, it performs clean feature fusion followed by detector-only bottom-up aggregation. A training-only enhancement head estimates background light, transmission, and clean radiance and enforces the Jaffe–McGlamery (JM) model via self-supervised reconstruction and guided refinement. This head is removed during inference. On the RUOD, URPC2022, and RUIE-UHTS datasets, UEAOD improves mean average precision (AP) by 2%–3% over a strong YOLOv12s baseline while preserving real-time throughput with a compact model size. These results indicate that explicit base–detail decomposition, coupled with physics-consistent supervision, yields domain-stable features and accurate, efficient underwater detection without incurring additional inference overhead. The source code is available at https://github.com/zhuyy116/UEAOD