An underwater object detection method integrating multi-dimensional attention and task-decoupling mechanism

Abstract

Underwater object detection is of significant practical importance for marine resource exploration, underwater robotic navigation, and marine ecological monitoring. However, underwater images are often severely degraded by light attenuation and scattering, suspended particulates, and complex background interference. These factors lead to low contrast, strong color distortion, and blurred object boundaries, which collectively pose substantial challenges to reliable object detection. To address these issues, this paper proposes an object detection framework specifically designed for underwater environments. First, we develop a dynamic image feature fusion backbone module, termed DyC2F (dynamic lightweight convolution to fusion). By embedding dynamic convolution into a dual-path architecture, DyC2F enhances the modeling of local textures and weak boundary features while maintaining low computational complexity. Second, a three-dimensional attention fusion mechanism, referred to as TDAF (three-dimensional attention fusion), is introduced to adaptively enhance multi-scale underwater features across channel, scale, and spatial dimensions, thereby improving the detectability of low-contrast small objects. Finally, the detection head is further optimized by incorporating a dynamic activation function, a task-decoupled alignment module, and a quality focal loss, which effectively mitigates feature conflicts between classification and localization in complex underwater backgrounds and improves prediction stability. Experimental results on representative underwater datasets demonstrate that the proposed method consistently outperforms mainstream approaches in terms of detection accuracy, robustness, and small-object recognition. Moreover, the proposed framework achieves a processing speed of 188.7 frames per second (fps), satisfying real-time detection requirements in complex underwater scenarios.