HAD-Net: A Hybrid Attention Driven Dynamic Network for AUV-Based Underwater Object Detection

Underwater object detection (UOD) is critical for enabling autonomous underwater vehicles (AUVs) to carry out underwater operations incomplex marine environments. However, the underwater domain poses severe challenges for visual perception due to low visibility, color distortion, and dynamic noise. To address these issues, Hybrid Attention Driven Dynamic Network (HAD-Net), which integrated a novel dynamic convolution module named FDConv (Full-Dimensional Convolution) was proposed. Unlike existing dynamic convolution approaches that only adapt over the kernel number dimension, FDConv introduced a four-way attention mechanism across the channel, filter, spatial, and kernel dimensions, enabling precise and adaptive feature modulation. By embedding FDConv into a YOLOv11-based detection framework, HAD-Net achieved superior performance in challenging underwater scenarios. Extensive experiments on the URPC2020 and TrashCan datasets demonstrated that HAD-Net outperforms existing detectors in both accuracy and robustness, while maintaining real-time inference speed. Our method provides a lightweight and deployable solution for perception tasks for autonomous underwater vehicles (AUVs).