利用改进Faster-RCNN识别小麦条锈病和黄矮病  

Recognizing stripe rust and yellow dwarf of wheat using improved Faster-RCNN

作　　者：毛锐[1] 张宇晨 王泽玺 高圣昌 祝涛 王美丽[1] 胡小平[2,3] Mao Rui;Zhang Yuchen;Wang Zexi;Gao Shengchang;Zhu Tao;Wang Meili;Hu Xiaoping(College of Information Engineering,Northwest A&F University,Yangling 712100,China;College of Plant Protection,Northwest A&F University,Yangling 712100,China;Key Laboratory of Integrated Pest Management on Crops in Northwestern Loess Plateau,Ministry of Agriculture and Rural Affairs,Yangling 712100,China)

机构地区：[1]西北农林科技大学信息工程学院,杨凌712100 [2]西北农林科技大学植物保护学院,杨凌712100 [3]农业农村部黄土高原作物有害生物综合治理重点实验室,杨凌712100

出　　处：《农业工程学报》2022年第17期176-185,共10页Transactions of the Chinese Society of Agricultural Engineering

基　　金：陕西省科技厅区域创新能力引导计划(2022QFY11-03);国家现代农业(小麦)产业技术体系项目(CARS-03-37);农业农村部农作物病虫鼠害疫情监测与防治项目;大学生创新训练项目(X202110712436)。

摘　　要：条锈病和黄矮病是严重威胁小麦生产的重大病害,病害的早期识别对病害防控具有重要意义。现有病害识别模型对相似表型症状识别困难,对早期病害的识别准确度低。为此,该研究构建了一种改进的快速区域卷积神经网络(Faster Regions with CNN Features,Faster-RCNN)的病害识别方法。该方法采用卷积核拆解和下采样延迟策略优化了深度残差网络(Deep Residual Neural Network,ResNet-50),用优化后的ResNet-50作为主干特征提取网络以增强所提取特征的表达力,同时简化模型的参数;并采用ROI(Region of Interest)Align改进ROI迟化层以降低特征量化误差,提升识别的精度。在自建的涵盖200余种不同发病时期、不同抗感性的小麦叶部图像数据集上进行试验,结果表明:改进的Faster-RCNN识别方法比其他SSD(Single Shot Multi-Box Detector)、YOLO(You Only Look Once)和Faster-RCNN网络模型的平均精度均值(mean Average Precision,m AP)分别提升了9.26个百分点、7.64个百分点和14.97个百分点。对小麦条锈病、黄矮病、健康小麦和其他黄化症状小麦识别的平均精度均值可达98.74%;对小麦条锈病和黄矮病轻、重症识别的平均精度均值可达91.06%。同时,模型损失函数值降低更快,整体性能表现更优。进一步开发小麦病害智能识别系统部署研究模型,使用微信小程序进行田间小麦病害的识别。在最大并发100的条件下,小程序平均返回时延为5.02 s,识别返回成功率为97.85%,对两种小麦病害及其细分轻重症识别的平均准确率为93.56%,能够有效满足实际应用需求,可用于指导病害的科学防控。Wheat stripe rust and wheat yellow dwarf have posed a great threat to the yield and quality of wheat.An accurate identification has important implications for the prevention and control of wheat diseases.However,the phenotypic symptoms are similar to the infected leaves of wheat stripe rust and wheat yellow dwarf.Particularly,drought,nutrient deficiency,and bacterial disease can lead to the chlorosis and yellowing of plant leaves.In addition,the infected leaves are also similar to the healthy ones,due to the indistinct phenotypic symptoms in the early stage of diseases.It is difficult to quickly and accurately distinguish them by the existing identification.In this study,an improved Faster Regions with CNN Features(Faster-RCNN)was proposed for disease identification.There were two improvement strategies.Firstly,three 3×3 grouping convolution and down-sampling delays were employed to optimize the Deep Residual Neural Network(ResNet-50),which was designed as the backbone feature extraction network,in order to enhance the fine feature extraction of the entire network.Secondly,the region of interest(ROI)alignment was employed to reduce the feature error caused by double quantization,instead of ROI pooling.As such,the subtle differences were recognized after alignment.Transfer learning was selected to improve the training speed of the model.The data augmentation was then utilized to reduce the over-fitting,in order to further enhance the recognition performance and generalization ability.The image data set of disease leaf was collected from more than 200 wheat varieties with different resistance and susceptibility to the diseases,covering various symptoms at different disease stages.A series of experiments were carried out to evaluate the improved strategy.The performance indicators were selected to verify the model,such as loss function convergence curve and mean average precision(mAP).The experimental results showed that the mAP of the improved Faster-RCNN reached 98.74%for the wheat stripe rust and wheat yellow dw

关 键 词：模型 病害识别 Faster-RCNN ResNet 分组卷积 数据增强 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] S512.1[自动化与计算机技术—计算机科学与技术]