Underwater object detection in riverine, lacustrine, and urban shoreline engineering environments is fundamentally constrained by degraded visual features and severe scale variation. These challenges cause small and low-contrast objects to be easily suppressed during multi-scale feature fusion and task prediction. To address these issues, this paper proposes YOLO-MFD(Multi-Scale Feature and Dynamic Head), a novel framework that jointly optimizes feature extraction, multi-scale fusion, and detection head design. First, we introduce a Cross-scale Unified Multi-scale Attention Network (CUMANet) to enhance contextual modeling and suppress environmental noise. Second, an Adaptive Feature Modulation (AFM) module is designed to achieve balanced multi-scale fusion and preserve shallow features, which are vital for detecting small and blurred underwater targets commonly encountered in complex shoreline waters. Third, a Dual-Pooling and Normalized Dynamic Head (DPNDyHead) is proposed to dynamically align and refine features, effectively mitigating task misalignment. Extensive experiments validate the robustness of the proposed framework. Compared with state-of-the-art detectors, YOLO-MFD achieves superior performance on public benchmark datasets, with notable improvements in mean Average Precision and recall metrics. Ablation studies further demonstrate the necessity and contribution of each proposed module to the overall detection performance in practical river and urban shoreline underwater inspection scenarios.
Gang et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: