摘要
为提高梨叶片病害发生程度诊断的效率和准确性,本文提出基于全局上下文级联R-CNN网络(Global context Cascade R-CNN,GC-Cascade R-CNN)的梨叶病斑计数方法。模型的主干特征提取网络嵌入全局上下文模块(Global context feature model,GC-Model),建立有效的长距离和通道依赖,增强目标特征信息。引入特征金字塔网络(Feature pyramid network,FPN)融合浅层细节特征和深层丰富语义特征。使用ROI Align替换ROI Pooling进行区域特征聚集,增强目标特征表达。最后利用多层级联网络对目标区域进行边框回归和分类,完成病斑计数任务。在梨叶病斑图像测试中,模型的各类病斑平均精确率均值(Mean average precision,mAP)达89.4%,检测单幅图像平均耗时为0.347 s。结果表明,模型能够有效地从梨叶片病害图像中检测出多类病斑目标,尤其对叶片炭疽病斑检测效果提升显著;不同种类梨叶片病害病斑计数值与真实值回归实验决定系数R^(2)均大于0.92,表明模型病斑计数准确率较高。
In order to improve the efficiency and accuracy of pear leaf disease degree diagnosis,a pear leaf disease spot counting method was proposed based on global context Cascade region-based convolutional neural network(GC-Cascade R-CNN).The backbone feature extraction network of the model was embedded in a global context feature model(GC-Model),to establish effective long-range dependency and channel dependency for enhancing the feature information.The model fused shallow detail features and deep rich semantic features by feature pyramid networks(FPN).ROI Align was used to replace ROI Pooling for regional feature aggregation and enhance the target feature representation.Bounding box regression and classification of target regions were performed by using multilayer Cascade networks to complete the disease spot counting task.In the test of pear leaf disease images,the mean average precision(mAP)of the model reached 89.4%for all types of disease spots,and a single image processing average time of 0.347 s,ensuring real-time operation while improving detection accuracy.The results showed that the model could effectively detect multiple types of disease spot targets from pear leaf disease images,especially for the detection of anthracnose spots;and the coefficient of determination R^(2) of the regression of disease spot counting values and true values of different kinds of pear leaf diseases were all greater than 0.92,indicating that the model had high accuracy of disease spot counting.This study solved the difficulty of pear leaves disease degree diagnosis,and provided a new idea for the diagnosis of pear disease conditions and symptoms in automated agricultural production.
作者
薛卫
程润华
康亚龙
黄新忠
徐阳春
董彩霞
XUE Wei;CHENG Runhua;KANG Yalong;HUANG Xinzhong;XU Yangchun;DONG Caixia(College of Artificial Intelligence,Nanjing Agricultural University,Nanjing 210095,China;College of Resources and Environmental Sciences,Nanjing Agricultural University,Nanjing 210095,China;Fruit Research Institute,Fujian Academy of Agriculture Sciences,Fuzhou 350013,China)
出处
《农业机械学报》
EI
CAS
CSCD
北大核心
2022年第5期237-245,共9页
Transactions of the Chinese Society for Agricultural Machinery
基金
财政部和农业农村部:国家现代农业产业技术体系项目(CARS28)。
关键词
梨树叶片
病斑计数
级联网络
全局上下文特征
注意力机制
小目标检测
pear leaf
disease spot counting
Cascade R-CNN
global context feature
attention mechanism
small object detection
