<p>Although target detection models have significantly improved in land-based scenarios, their generalization in complex underwater environments remains severely constrained by unique sources of interference, such as dynamic physical noise, dense target occlusion, and morphological variability induced by scale changes and irregular contours of marine organisms. These factors collectively limit the perceptual accuracy and robustness of existing models. To address these challenges, we propose AquaFuse-Net. The model first introduces the hierarchical dilated residual module (HDR) to address insufficient multi-scale feature fusion. This is achieved through hierarchical convolutional design and structural reparameterization, enhancing feature discrimination in low-visibility environments. Second, we present the BiPath dual aggregation module (BPDA), which effectively suppresses the impact of dynamic physical noise on feature expression. To counter the loss of fine-grained details, the DySample module optimizes the feature pyramid upsampling strategy, improving recovery of small and occluded targets. Furthermore, we introduce the MPDIoU loss function to better accommodate irregular contours of underwater organisms. Experimental results demonstrate that AquaFuse-Net achieves 83.84% mAP on UTDAC2020. It also shows robust performance on URPC2020 (82.80%) and RUOD (84.38%). These results confirm the model’s strong generalization capability across extreme underwater scenarios. The framework’s high inference speed, efficiency, and parallelizable design make it suitable for real-time deployment on autonomous underwater vehicles (AUVs), underscoring its direct relevance to the field of HPC-enabled marine robotics.</p>