基于YOLOv5的水面垃圾旋转目标检测模型
Rotating Target Detection Model of Water Surface Garbage Based on YOLOv5

水面垃圾污染给生态环境带来巨大威胁，无人船抓拣为水面垃圾清理提供了一种更为高效环保的方法。水面垃圾目标检测是无人船研究的关键技术之一。传统的水平目标检测在环境复杂，目标角度不规则的情况下检测精度较低且抓拣失误率较高。针对此问题，本文提出一种基于改进YOLOv5的水面垃圾旋转目标检测模型。该模型优化了YOLOv5模型的检测头，可以对任意角度的水面垃圾生成定向检测框。在此基础上，本文进一步从三个方面改进了YOLOv5在水面垃圾检测中的性能：提出动态平滑角度损失函数以增强角度预测能力；通过增强浅层特征并引入注意力机制来优化骨干网络；改进特征聚合网络，增强特征多尺度融合效果。实验结果表明，该方法对水面垃圾的目标检测性能优于其它深度学习方法。
Water surface garbage pollution poses a significant threat to the ecological environment, and unmanned boat grabbing offers a more efficient and environmentally friendly method for water surface garbage cleanup. In the field of unmanned boat research, the precise detection of water surface garbage is one of the core technologies. Traditional horizontal target detection has lower detection accuracy and a higher rate of grabbing errors in complex environments with irregular target angles. To address this issue, this paper proposes a water surface garbage rotating target detection model based on an improved YOLOv5. The model optimizes the detection head of the YOLOv5 model to generate oriented detection boxes for water surface garbage at any angle. On this basis, the paper further improves the performance of YOLOv5 in water surface garbage detection from three aspects: proposing a dynamic smooth angle loss function to enhance angle prediction capability; optimizing the backbone network by enhancing shallow features and introducing attention mechanisms; and improving the feature aggregation network to enhance the effect of multiscale feature fusion. Experimental results show that this method outperforms other deep learning methods in the target detection performance of water surface garbage.