摘要
为提高果树病虫害危害程度分级精度进而更好地指导果园病虫害防治,采用迁移学习技术与GoogLeNet模型相结合的方法,对6种果园作物的25类病虫害样本进行识别与危害程度分级研究;同时,探究不同数据集大小以及不同优化算法对模型性能的影响;基于MATLAB平台设计了一款可视化的病虫害识别与分级系统。结果表明:1)基于迁移学习的GoogLeNet模型,对病虫害识别精度可达99.35%,危害程度分级精度可达92.78%;2)在相同训练参数下,本研究模型比AlexNet、VGG-16、ResNet-18、SqueezeNet、原GoogLeNet及MobileNet-v2模型验证精度提高了2.38%~11.44%,并且收敛速度最快;3)本研究模型识别精度随着数据集的增大而提高;在3种优化算法中SGDM算法耗时最短且精度最高,更适合本研究模型。通过拍摄果树叶片病害区域图像,本研究设计的系统能够在0.43s左右准确识别出果树种类、病害类型以及危害等级等信息。
To improve the grading accuracy of fruit tree diseases and pests and better guide the prevention and control of orchard diseases and pests,the Transfer Learning technology and GoogLeNet model were combined to recognize and grade the damage degree of 25 kinds of pests and diseases from 6 kinds of orchard crops.At the same time,the impact of different data set sizes and optimization algorithms on model performance was studied.A visual pest recognition and grading system was designed based on the MATLAB platform.The results showed that:1)The GoogLeNet model based on Transfer Learning had an up to 99.35%accuracy in recognizing pests and diseases,and an accuracy of 92.78%in the grade of damage;2)Under the same training parameters,the verification accuracy of this research model was 2.38%to 11.44%higher than that of AlexNet,VGG-16,ResNet-18,SqueezeNet,the original GoogLeNet and MobileNet-v2 models,and the convergence speed was the fastest;3)The recognition accuracy of this research model enhanced with the increase of the data set;Among the three optimization algorithms,the SGDM algorithm took the shortest time and had the highest accuracy,which was more suitable for this research model.By taking images of the diseased area of fruit tree leaves,the system designed in this study can accurately recognize the name of fruit trees,type of diseases and grade of damage within 0.43 s.
作者
万军杰
祁力钧
卢中奥
周佳蕊
张豪
WAN Junjie;QI Lijun;LU Zhongao;ZHOU Jiarui;ZHANG Hao(College of Engineering,China Agricultural University,Beijing 100083,China)
出处
《中国农业大学学报》
CAS
CSCD
北大核心
2021年第11期209-221,共13页
Journal of China Agricultural University
基金
国家重点研发计划(2017YFD0701400,2016YFD0200700)。
