This paper proposes a dynamic YOLO detector for small underwater object detection, addressing the challenges posed by small objects in underwater environments. The proposed method includes a lightweight backbone network based on deformable convolution v3, a unified feature fusion framework with channel-, scale-, and spatial-aware attention mechanisms, and an extended decoupled head to disentangle and align classification and localization tasks. The backbone network is designed to efficiently extract features from small objects, while the feature fusion framework enhances the detection of small objects by dynamically combining features from different scales. The extended decoupled head improves the alignment of classification and localization tasks, leading to better detection performance. The proposed dynamic YOLO is evaluated on benchmark datasets, including the DUO dataset, Pascal VOC, and MS COCO, demonstrating superior performance. On the DUO dataset, dynamic YOLO outperforms recent state-of-the-art methods by +0.8 AP and +1.8 AP$_{S}$. Ablation studies validate the effectiveness and efficiency of each design in dynamic YOLO. The results show that the proposed method achieves state-of-the-art performance in detecting small underwater objects.This paper proposes a dynamic YOLO detector for small underwater object detection, addressing the challenges posed by small objects in underwater environments. The proposed method includes a lightweight backbone network based on deformable convolution v3, a unified feature fusion framework with channel-, scale-, and spatial-aware attention mechanisms, and an extended decoupled head to disentangle and align classification and localization tasks. The backbone network is designed to efficiently extract features from small objects, while the feature fusion framework enhances the detection of small objects by dynamically combining features from different scales. The extended decoupled head improves the alignment of classification and localization tasks, leading to better detection performance. The proposed dynamic YOLO is evaluated on benchmark datasets, including the DUO dataset, Pascal VOC, and MS COCO, demonstrating superior performance. On the DUO dataset, dynamic YOLO outperforms recent state-of-the-art methods by +0.8 AP and +1.8 AP$_{S}$. Ablation studies validate the effectiveness and efficiency of each design in dynamic YOLO. The results show that the proposed method achieves state-of-the-art performance in detecting small underwater objects.