基于多尺度和注意力机制的番茄病害识别方法  

Tomato disease recognition scheme based on multi-scale and attention mechanism

作　　者：张宁 吴华瑞[2,3,4] 韩笑 缪祎晟 ZHANG Ning;WU Huarui;HAN Xiao;MIAO Yisheng(College of Computer and Information Engineering, Beijing University of Agriculture, Beijing 102206, China;National Engineering Research Center for Information Technology in Agriculture, Beijing 100097, China;Beijing Research Center for Information Technology in Agriculture, Beijing 100097, China;Key Laboratory of Agri-Informatics, Ministry of Agriculture and Rural Affairs, Beijing 100097, China)

机构地区：[1]北京农学院计算机与信息工程学院,北京102206 [2]国家农业信息化工程技术研究中心,北京100097 [3]北京农业信息技术研究中心,北京100097 [4]农业农村部农业信息技术重点实验室,北京100097

出　　处：《浙江农业学报》2021年第7期1329-1338,共10页Acta Agriculturae Zhejiangensis

基　　金：国家自然科学基金(61871041);国家大宗蔬菜产业技术体系岗位专家项目(CARS-23-C06);石家庄市科学技术研究与发展项目(201490074A)。

摘　　要：番茄病害的及时发现与治理有助于提高番茄产量与质量,增加农户经济收益。利用物联网和人工智能可以无损害有效检测番茄病害,该研究提出了一种改进的AT-InceptionV3(Attention-InceptionV3)神经网络番茄叶部病害检测模型,该网络以InceptionV3为主干网络,结合多尺度卷积和注意力机制CBAM(convolutional block attention module,CBAM)模块,增强了病害信息表达并抑制无关信息干扰;同时引入迁移学习,防止样本数据量较少时出现过拟合的情况。为了评价优化模型的有效性,在Plant Village公开番茄病害数据集上进行了实验仿真测试。改进的模型在测试阶段对番茄健康叶片、细菌性斑疹病、晚疫病、叶霉病和黄曲病5种番茄常见叶片图像分类准确率达到98.4%,优化效果显著。为了进一步验证该方法在不同物联网中的普适性,实验对比了模型对不同分辨率病害图像的分类效果,结果表明,图像精度部分损失不会降低病害分类准确率。该模型能够为番茄温室智能网络决策判断提供重要依据。The timely detection and management of tomato diseases can help increase the yield and quality of tomatoes and increase the economic benefits of farmers.The internet of things and artificial intelligence can effectively detect tomato diseases without damage.This paper proposed an improved AT-InceptionV3(Attention-InceptionV3)neural network tomato leaf disease detection model.The network used InceptionV3 as the backbone network,combined with multi-scale convolution and attention mechanism CBAM(convolutional block attention module,CBAM)module,which enhanced disease information expression and suppressed irrelevant information interference.At the same time,transfer learning was introduced to prevent over-fitting when the sample data volume was small.In order to evaluate the effectiveness of the optimization model,an experimental simulation test was conducted on the public tomato disease data set of Plant Village.The improved model had an accuracy of 98.4%in classification of five leaf images of healthy tomato leaves,bacterial spot disease,late blight,leaf mold and yellow leaf curl vircus in the test stage,and the optimization effect was significant.In order to further verify the universality of this method in different Internet of Things,experiments had compared the model's classification effect on disease images with different resolutions.The results showed that partial loss of image resolution would not reduce the accuracy of disease classification.This model could be used for tomato greenhouse intelligence and provide an important basis for network decision-making and judgment.

关 键 词：番茄 多尺度卷积 注意力机制 迁移学习 病害识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] S436.412[自动化与计算机技术—计算机科学与技术]