基于深度残差网络的玉米病害识别  

Corn disease recognition based on deep residual network

作　　者：刘翱宇 吴云志[1,2] 朱小宁 范国华 乐毅[1,2] 张友华 LIU Ao-yu;WU Yun-zhi;ZHU Xiao-ning;FAN Guo-hua;YUE Yi;ZHANG You-hua(School of Information and Computer Science, Anhui Agricultural University, Hefei 230036, China;Anhui Provincial Engineering Laboratory for Beidou Precision Agriculture Information, Hefei 230036, China;School of Tea and Food Technology, Anhui Agricultural University, Hefei 230036, China)

机构地区：[1]安徽农业大学信息与计算机学院,安徽合肥230036 [2]安徽省北斗精准农业信息工程实验室,安徽合肥230036 [3]安徽农业大学茶与食品科技学院,安徽合肥230036

出　　处：《江苏农业学报》2021年第1期67-74,共8页Jiangsu Journal of Agricultural Sciences

基　　金：国家重点研发计划项目(2017YFD0301303);安徽省科技重大专项(18030901029);安徽省高校自然科学研究项目(KJ2019A0211);安徽省大学生创新创业训练计划项目(S202010364210)。

摘　　要：针对人工诊断玉米病害成本高、效率低、时延长等问题,提出一种基于深度残差网络的玉米病害识别网络TFL-ResNet。TFL-ResNet网络基于ResNet50网络,首先引入Focal Loss损失函数使模型专注于难分类的病害样本,其次将ResNet50网络在PlantVillage数据集训练好的参数迁移到改进网络上以完成构建。采用的玉米病害数据集涉及健康植株、大斑病、灰斑病、锈病4种标签,并使用旋转、翻转、平移等操作对数据集进行数据增强与扩充。对数据集进行训练和测试,与VGG16等对照模型相比,TFL-ResNet网络收敛速度更快、分类效果更好,平均识别准确率高达98.96%。通过观察精准率、召回率、混淆矩阵等评价指标得出TFL-ResNet网络具有较好的鲁棒性和泛化能力,可用于玉米病害智能诊断。Aiming at the problems of high cost,low efficiency,time extension and so on for manual diagnosis of corn diseases,a corn disease identification network TFL-ResNet based on deep residual network was proposed.The TFL-ResNet was based on the ResNet50.Firstly,the focal loss function was introduced to make the model focus on disease samples which were difficult to classify.Secondly,the parameters of ResNet50 trained in PlantVillage data set were migrated to the improved network to complete the construction.The corn disease data set used in this study involved four labels:healthy plants,leaf blight,gray leaf spot and rust.Rotation,flip,translation and other operations were used to enhance and expand the data set.Compared with VGG16 and other control models,TFL-ResNet had faster convergence speed and better classification effect.The average recognition accuracy of TFL-ResNet was 98.96%.By observing the model evaluation indicators such as precision rate,recall rate,confusion matrix,TFL-ResNet has good robustness and generalization ability,which can be used for intelligent diagnosis of corn diseases.

关 键 词：病害图像识别 深度残差网络 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]