Novel Dynamic Feature Fusion Stragegy for Detection of Small Underwater Marine Object

Underwater marine object detection has been shown to be significant in the community of marine science and engineering, and has been extensively employed in many practical applications, including commercial fisheries, marine exploration, etc. A typical small object detection trap frequently exists due to the fact that the majority of objects of interest in the underwater environment are usually small. Detection of small marine object is highly desired. This paper presents a novel dynamic feature fusion algorithm based on spatial-aware and scale-aware attention that adaptively fuses different scale feature maps for the detection of small underwater marine objects. Leveraging the proposed dynamic feature fusion algorithm, an improved YOLOX-DFF model is proposed to detect marine organisms in natural seabed environments. To alleviate conflicts between classification and regression tasks in the detection head, the decoupled head of YOLOX-DFF is also augmented with a task-aware attention module. The proposed model surpasses state-of-the-art methods by a large margin of $+0.7mAP$, and $+2.3AP_{{S}}$ on the DUO dataset. The effectiveness and efficiency of each module are further demonstrated by the ablation studies on PASCAL VOC 0712 datasets, achieving excellent performance with $+1.3mAP$, and $+3.0AP_{S}$ improvements.