摘要
[目的]本文旨在探索实时监测温室虫情和精准防控虫害的方法。[方法]设计了一种基于诱虫板图像背景均匀化的自适应分割方法,结合基于随机森林(random forest,RF)的图像识别算法识别4类温室番茄害虫(烟粉虱、潜叶蝇、果蝇和蚜虫)并计数。该方法首先提取诱虫板图像RGB(red-green-blue)颜色模型B分量和HSV(hue-saturation-value)颜色模型V分量,然后分别对2张图像分段调整背景灰度值得到均匀背景诱虫板灰度图像,再利用最大类间方差法确定阈值分割图像,经形态学处理后融合2张诱虫板二值图像,最后提取害虫区域的6个颜色特征、8个形状特征和6个纹理特征,训练随机森林以识别害虫并计数。[结果]对比分析Sauvola局部阈值法、Prewitt边缘分割法、k-means聚类法以及本文设计的自适应分割方法,结果表明基于背景均匀化的自适应分割方法效果最好,平均分割准确率为95.34%。对比分析7种特征向量组合下随机森林、C-SVC(C-support vector classification)和BP(back propagation)神经网络3种分类方法,结果表明综合颜色特征向量、形状特征向量和纹理特征向量作为输入的随机森林算法识别效果更好,对烟粉虱、潜叶蝇、果蝇和蚜虫的识别准确率分别为93.89%、90.71%、91.54%和90.40%。[结论]本文设计的方法能够实现诱虫板上4类害虫的识别和计数,可以为温室虫情监测与预警提供参考。
[Objectives]The paper aimed to explore the method of real-time monitoring of insect pest situation in greenhouse and accurate control of the attacks by pests.[Methods]In this paper,an adaptive segmentation method based on background homogenization,combined with a random forest(RF)classification method was designed to identify and count four categories of tomato pests(Bemisia tabaci,Chromatomyia horticola,Drosophila melanogaster and Laingia psammae)in greenhouse.Firstly,this method extracted the B components from RGB(red-green-blue)color model and the V components from HSV(hue-saturation-value)color model in trapping board image,and then it adjusted the background gray level in sections to achieve the uniform background trapping board image respectively.Then it used the maximum inter-class variance method to determine the threshold to segment the image,and two binary images of trapping board were fused after morphological processing.Finally,6 color features,8 shape features and 6 texture features extracted from the pest region were inputted into the random forest classifier,which was trained to identify and count the pests.[Results]By comparing and analyzing Sauvola local threshold method,Prewitt edge segmentation method,k-means clustering method and the method of this paper,it was found that the adaptive segmentation based on background homogenization had the best effects,and the average segmentation accuracy was 95.34%.By comparing and analyzing the method of C-SVC(C-support vector classification),BP(back propagation)neural network,random forest with 7 kinds of feature vectors,it was found that random forest classification method with the inputs of the integration of the color features,the shape features,and the texture features was more effective,and the respective recognition accuracy for B.tabaci,C.horticola,D.melanogaster and L.psammae was 93.89%,90.71%,91.54%and 90.40%.[Conclusions]The algorithm can realize the recognition and counting of 4 kinds of pests on the trapping board and provide theoretical and me
作者
卜俊怡
孙国祥
王迎旭
魏天翔
汪小旵
BU Junyi;SUN Guoxiang;WANG Yingxu;WEI Tianxiang;WANG Xiaochan(College of Engineering/Jiangsu Modern Facility Agricultural Technology and Equipment Engineering Laboratory,Nanjing Agricultural University,Nanjing 210031,China)
出处
《南京农业大学学报》
CAS
CSCD
北大核心
2021年第2期373-383,共11页
Journal of Nanjing Agricultural University
基金
国家重点研发计划项目子课题(2019YFD1001902-11)。
关键词
诱虫板图像
背景均匀化
图像分割
害虫识别
随机森林
trapping board image
background homogenization
image segmentation
pest recognition
random forest
