摘要
针对河道“四乱”检测图像中目标尺寸小、河道背景复杂等造成目标难以识别等问题,提出一种YOLO v8n-SPE-SL(Small SPD-Conv-ECA SiLuan)模型,旨在实现河道“四乱”目标的快速精准识别。通过增加小目标检测层,显著提升模型对小尺度目标的特征提取能力;通过引入SPD-Conv模块替换原模型中步长为2的部分卷积,减少细节信息的损失;通过在部分C2f模块当中加入ECA(Efficient channel attention)注意力机制,显著提高模型在复杂背景下的目标识别能力。在此基础上设计了河道“四乱”巡查系统。基于构建的数据集对改进模型进行实验对比,结果表明,YOLO v8n-SPE-SL模型的精确率、召回率和平均精度均值分别达到96.3%、91.9%和95.7%,与YOLO v8n模型相比,分别提升1、2.5、1.6个百分点;其中引入小目标检测层使mAP@50提升0.7个百分点,引入SPD-Conv模块使误检率降低23.6%,引入ECA机制将mAP@50-95提高2.7个百分点。巡查系统可用于实现“四乱”目标的精准识别与展示,助力幸福河湖建设。
In order to solve the problems of small target size and complex background in the river channel“four chaos”detection image,an improved YOLO v8n-SPE-SL(Small SPD-Conv-ECA SiLuan)model was proposed to quickly and accurately identify the“four chaos”targets in the river channel.By adding a small target detection layer,the difficult problem of small target recognition in the river was solved.By introducing the SPD-Conv module to replace the partial convolution with step size of 2 in the original model,the loss of detail information was reduced.By adding the efficient channel attention(ECA)mechanism to some C2f modules,the ability to recognize small targets of the“four chaos”in the river was improved,and on this basis,the“four chaos”patrol system of the river was designed.Based on the constructed dataset,the results showed that the average accuracy,recall rate and average accuracy of the YOLO v8n-SPE-SL model reached 96.3%,91.9%and 95.7%,which were improved by 1,2.5 and 1.6 percentage points respectively compared with that of the YOLO v8n model.The introduction of the small target detection layer improved the mAP@50 by 0.7 percentage points,the SPD-Conv module reduced the false detection rate by 23.6%,and the ECA mechanism increased the mAP@50-95 by 2.7 percentage points.The inspection system can be used to achieve precise identification and display of the four“chaos”phenomena(“unauthorized occupation,”“illegal construction,”“random piling,”and“illegal mining”)within river management areas,contributing to the construction of happy rivers and lakes.
作者
刘玲
马晓艳
孙天玥
申孝军
刘冬梅
苏昌发
LIU Ling;MA Xiaoyan;SUN Tianyue;SHEN Xiaojun;LIU Dongmei;SU Changfa(College of Computer and Information Engineering,Tianjin Agricultural University,Tianjin 300392,China;College of Water Conservancy Engineering,Tianjin Agricultural University,Tianjin 300392,China;The Water Affairs Bureau of Jizhou District,Tianjin,Tianjin 301900,China)
出处
《农业机械学报》
北大核心
2025年第8期163-171,共9页
Transactions of the Chinese Society for Agricultural Machinery
基金
天津市人社局项目+团队重点培养专项(XB202016)。
