基于改进YOLOv5的内河道船舶检测方法研究
Research on ship detection on inland waterways with improved YOLOv5
丁飞 1张祥林 1刘明君1
作者信息
- 1. 阜阳师范大学 计算机与信息工程学院,安徽 阜阳 236037
- 折叠
摘要
为解决内河航道通航时船舶相互遮挡导致的错检和漏检问题,提出一种基于改进 YOLOv5 模型的内河道船舶检测方法.该方法采用C2f模块以捕捉和融合多尺度特征,增强低级特征的语义信息;引入Shuffle Attention模块强化特征表示,使模型能够聚焦于信息量更大的区域,并有效抑制无关特征;同时,采用Wise-IoU损失函数,有效防止低质量锚框产生有害梯度,加速模型的优化过程.研究结果表明:改进后模型的平均精度mAP@0.5达到 98.9%,mAP@0.5:0.95 达到 79.1%,较原YOLOv5 模型分别提高了 1.2 和 2.8 个百分点.此外,针对数据集中的 6 种船舶分别进行实验,实验结果显示各种船舶的检测精度均有提升,其中内河航道常见的矿砂船的mAP@0.5提升了 1.5 个百分点,mAP@0.5:0.95 提升5.9 个百分点.研究结论为内河道船舶检测提供了可靠的技术支持.
Abstract
To address the issue of misdetection and missed detection caused by vessel occlusion in inland waterway navigation,this study proposes an inland waterway vessel detection method based on an improved YOLOv5 model.The method incorporates the C2f module to capture and integrate multi-scale features,enhancing the smantic information of low-level features.Additionally,the Shuffle Attention module is introduced to strengthen feature representation,enabling the model to focus on more informative regions while effectively suppressing irelevant features.The Wise-IoU loss function is also employed to prevent low-quality anchor boxes from geneating harmful gradients,thereby accelerating the optimization process.Experimental results demonstrate that the improved model achieves a mean average precision(mAP@0.5)of 98.9%and an mAP@0.5:0.95 of 79.1%,representing an increase of 1.2 and 2.8 percentage points,respectively,compared to the original YOLOv5 model.Furthermore,experiments conducted on six types of vessels in the dataset reveal that detection accuracy has improved across all vessel types.Notably,for the commonly encountered sand carriers in inland waterways,the mAP@0.5 increased by 1.5 percentage points,and the mAP@0.5:0.95 improved by 5.9 percentage points.These findings provide reliable technical support for inland waterway vessel detection.
关键词
内河航道/船舶检测/YOLOv5模型/注意力机制/Wise-IoU损失函数Key words
inland waterway/ship inspection/YOLOv5 model/attention mechanism/Wise-IoU loss function引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
万方数据