YOLOv8n-PFA: a parallel fusion attention network for enhanced target detection in challenging environments

Introduction Underwater target detection plays a crucial role in marine environmental monitoring and ocean exploration. However, accurate detection remains challenging due to low illumination, blurred small objects, and complex background interference. Although convolutional neural network-based detectors have improved detection performance, many existing approaches are computationally expensive, limiting their deployment on resource-constrained underwater platforms. Methods To address these challenges, we propose YOLOv8n-PFA, a lightweight and high-precision underwater object detection framework. The proposed method introduces a novel Parallel Fusion Attention (PFA) module that models channel and spatial attention in parallel using residual connections to enhance discriminative features while suppressing background noise. The Wise Intersection over Union (WIoUv3) loss is incorporated to stabilize training and improve localization accuracy. Additionally, depth-wise convolutions (DWConv) are strategically applied to reduce model parameters and computational complexity. To further validate generalization capability, the PFA module is also integrated into YOLOv11n. Results Experimental results show that YOLOv8n-PFA achieves 84.2% mean Average Precision (mAP) on the URPC2020 dataset with 2.68 M parameters and 7.7 GFLOPs, and 84.8% mAP on the RUOD dataset with 2.98 M parameters and 7.9 GFLOPs. When integrated into YOLOv11n, the model achieves 84.7% mAP on URPC2020 and 85.3% on RUOD with only 2.76 M parameters and 6.5 GFLOPs. Across both datasets, the proposed approach improves mAP by 2.8-4.1% over baseline models while maintaining a lightweight architecture. Discussion The results demonstrate that the proposed framework provides an effective and computationally efficient solution for real-time underwater target detection in challenging marine environments. The consistent performance gains across different YOLO generations further confirm the scalability and robustness of the proposed PFA module.