This paper addresses the challenge of detecting small underwater objects, a critical issue in marine exploration. The authors propose a dynamic YOLO detector, which includes a lightweight backbone network based on deformable convolution v3, a unified feature fusion framework, and an extended decoupled head. The lightweight backbone network is designed to efficiently extract features from small objects, while the unified feature fusion framework integrates channel-wise, scale-wise, and spatial-aware attention mechanisms to enhance multi-scale feature fusion. The extended decoupled head disentangles and aligns classification and localization tasks to improve localization accuracy. Extensive experiments on benchmark datasets, including the DUO dataset, Pascal VOC, and MS COCO, demonstrate that the proposed dynamic YOLO detector outperforms state-of-the-art methods by a significant margin, achieving +0.8 AP and +1.8 AP$_S$ on the DUO dataset. Ablation studies further validate the effectiveness and efficiency of each component of the dynamic YOLO model.This paper addresses the challenge of detecting small underwater objects, a critical issue in marine exploration. The authors propose a dynamic YOLO detector, which includes a lightweight backbone network based on deformable convolution v3, a unified feature fusion framework, and an extended decoupled head. The lightweight backbone network is designed to efficiently extract features from small objects, while the unified feature fusion framework integrates channel-wise, scale-wise, and spatial-aware attention mechanisms to enhance multi-scale feature fusion. The extended decoupled head disentangles and aligns classification and localization tasks to improve localization accuracy. Extensive experiments on benchmark datasets, including the DUO dataset, Pascal VOC, and MS COCO, demonstrate that the proposed dynamic YOLO detector outperforms state-of-the-art methods by a significant margin, achieving +0.8 AP and +1.8 AP$_S$ on the DUO dataset. Ablation studies further validate the effectiveness and efficiency of each component of the dynamic YOLO model.