基于改进YOLOv5算法的河道水面漂浮物多视角目标检测  

Multi Perspective Object Detection of Floating Debris on River Surface Based on Improved YOLOv5 Algorithm

作　　者：金志浩 黄光团[1] Jin Zhihao;Huang Guangtuan(School of Resources and Environmental Engineering,East China University of Science and Technology,Shanghai 200237,China;Shanghai Hongkou District Environmental Monitoring Station,Shanghai 200083,China)

机构地区：[1]华东理工大学资源与环境工程学院,上海200237 [2]上海市虹口区环境监测站,上海200083

出　　处：《绿色科技》2025年第4期133-137,共5页Journal of Green Science and Technology

摘　　要：河道水面漂浮物是河道表观污染的重要源头,对其进行高效、精准的识别是改善水环境生态质量的重要途径。由于漂浮物具有场景复杂度高、移动频率高、形状不规则及多尺度形态变化等特点,传统的检测方法已难以满足复杂水面场景的大范围漂浮物识别需求,基于深度学习的YOLO图像识别算法为漂浮物智能识别与定位提供了新的技术支撑。为提升YOLOv5算法在小目标物和类别不均衡状态下检测能力,本文通过增加数据预处理模块和替换原有损失函数,提升其对小目标的检测性能。结果表明:相较于传统的YOLOv5算法,改进算法的平均精度均值(mAP)提高了8.86%,各类漂浮物的识别精度均有所提升。此外,混淆矩阵显示,该改进算法对于8种类型的漂浮物均有着较高的正确识别率。数据集的精确率-置信度曲线图(P曲线)和精确率-召回率曲线图(P-R曲线)显示,该改进算法对不同类型漂浮物均有较高的识别精确率,整体精确率达到97.0%。其中,船只的P-R曲线基本上“包住”其他类型漂浮物对应的曲线,其检测精度最高,与P曲线结果相一致。可视化结果显示:该算法在数码、固定和无人机摄像头3种识别视角下均具有较高的检测置信度和较低的漏检率,说明改进后的YOLOv5对目标学习得更充分,对重叠目标、小目标和特征模糊目标漏检的情况得到了有效的缓解。Floating debris on river surfaces is a significant source of apparent pollution in waterways,and its efficient and accurate identification is a crucial approach to improving the ecological quality of the water environment.Due to the high complexity of the scene,high frequency of movement,irregular shape,and multi-scale morphological changes of floating debris,traditional detection methods are no longer able to meet the large-scale floating debris recognition needs of complex water surface scenes.The YOLO image recognition algorithm based on deep learning provides new technical support for intelligent recognition and localization of floating debris.To enhance the detection capability of YOLOv5 algorithm in small target debris and imbalanced categories,this paper improves its detection performance for small targets by adding a data preprocessing module and replacing the original loss function.The results show that compared to the traditional YOLOv5 algorithm,the improved algorithm has increased the average accuracy mean(mAP)by 8.86%,and the recognition accuracy of various floating debris has been improved.In addition,the confusion matrix shows that the improved algorithm has a high correct recognition rate for 8 types of floating debris.The precision confidence curve(P-curve)and precision recall curve(P-R curve)of the dataset show that the improved algorithm has high recognition accuracy for different types of floating debris,with an overall accuracy of 97.0%.Among them,the P-R curve of the vessel basically"envelops"the curves corresponding to other types of floating debris,and its detection accuracy is the highest,consistent with the results of the P-curve.The visualization results show that the algorithm has high detection confidence and low missed detection rate in three recognition perspectives:digital,fixed,and drone cameras.This indicates that the improved YOLOv5 demonstrates more comprehensive learning for targets,effectively mitigating the issues of missed detection for overlapping targets,small targets,and

关 键 词：漂浮物 目标检测 YOLOv5 可视化 

分 类 号：X703[环境科学与工程—环境工程]