An underwater dual-modal denoising detection network for illumination fluctuation and dense occlusion scenes

Background Underwater object detection is a critical enabling technology for intelligent ocean exploration. However, light scattering and absorption in underwater environments cause significant RGB image degradation and texture loss, leading to weak foreground feature representation and limited detection accuracy in densely occluded scenes. Methods To address these challenges, this paper proposes the Underwater Dual-modal Detection Network (UW-DualDet), which incorporates depth information to compensate for RGB degradation and enhance detection performance. Two plug-and-play modules are designed: the Underwater Denoise Feature Fusion (UDFF) module suppresses dual-modal noise and adaptively exploits the complementarity between RGB texture and depth geometry to mitigate single-modality failure under extreme illumination, while the Underwater Feature Enhancement Module (UFEM) improves feature representation and multiscale semantic perception through multi-scale depthwise separable convolutions and a multi-dimensional weighting mechanism. Furthermore, a Gaussian-filterenhanced YOLO26 detection head (D26Head) is introduced to reduce missed and false detections in densely occluded scenarios by balancing positive sample coverage and localization accuracy through a dual-branch optimization strategy, while also suppressing feature noise. Results UW-DualDet achieves a mean Average Precision (mAP) of 86.7%, outperforming all compared state-of-the-art methods. It demonstrates superior robustness across three representative scenarios: 88.2% AP₅₀ in clear water, 79.6% AP₅₀ in turbid environments, and 75.3% AP₅₀ in low-light deep water. The model maintains an inference speed of 116 FPS, meeting real-time operational requirements. Discussion The proposed method effectively addresses the core challenges of underwater object detection by leveraging multimodal complementarity and task-specific optimization. It provides robust technical support for underwater intelligent operations and ecological monitoring. Future work will focus on improving depth estimation accuracy and extending the model to more extreme underwater environments.