基于深度学习的病害识别方法研究  

Research on Disease Recognition Method Based on Deep Learning

作　　者：王君婵[1] 洪俐 朱少龙 刘涛[2,3] 吴旭江 王慧[1] 孙成明[2,3] Wang Junchan;Hong Li;Zhu Shaolong;Liu Tao;Wu Xujiang;Wang Hui;Sun Chengming(Lixiahe Institute of Agricultural Sciences,Key Laboratory of Wheat Biology and Genetic Improvement for Low&Middle Yangtze Valley,Ministry of Agriculture and Rural Affairs,Yangzhou 225007,Jiangsu;Jiangsu Key Laboratory of Crop Genetics and Physiology,Jiangsu Key Laboratory of Crop Cultivation and Physiology,Agricultural College of Yangzhou University,Yangzhou 225009,Jiangsu;Jiangsu Co-Innovation Center for Modern Production Technology of Grain Crops,Yangzhou University,Yangzhou 225009,Jiangsu)

机构地区：[1]江苏里下河地区农业科学研究所/农业农村部长江中下游小麦生物学与遗传育种重点实验室,江苏扬州225007 [2]江苏省作物遗传生理重点实验室/江苏省作物栽培生理重点实验室/扬州大学农学院,江苏扬州225009 [3]江苏省粮食作物现代产业技术协同创新中心/扬州大学,江苏扬州225009

出　　处：《农业展望》2023年第8期90-99,共10页Agricultural Outlook

基　　金：江苏省农业科技自主创新资金(CX(21)3063);江苏省重点研发计划(现代农业)项目(BE2022335;BE2022338);江苏省省级农业科技创新与推广补助专项“江苏省省级作物种质资源库(里下河地区农作物)建设”。

摘　　要：当前在大面积生产中,主要病害的发生严重威胁农作物健康持续发展,亟须高效的病害识别方法来解决人工识别耗时、误判及效率低等问题。通过SONY@6300数码相机和开源数据库获取3种病害图像各150张,利用Labelme标注图片病斑,经过图像增强得到2250张图片并构成数据集,按照8∶1∶1的比例划分为训练集、验证集和测试集。训练Deeplabv3+、U-Net、U-Net++等3种神经网络模型,对3种病害的图片进行预测并评价,比较在相同迭代次数下的损失曲线;根据对测试集进行预测的混淆矩阵,计算其精准度、召回率和F1得分。综合对比这3种模型的性能,结果表明U-Net++效果最好,3个评价指标均在97%以上,对病害的分割性能也优于Deeplabv3+和U-Net模型。研究结果可为生产上病害防治提供科学指导,也可为其他作物的病害识别提供技术参考。In large-scale production, the occurrence of major diseases seriously threatens the healthy and sustainable development of the crop industry, an efficient disease identification method is urgently needed to solve the problems of time-consuming, misjudgment and low efficiency of manual identification.In this study, 150 images for each of the three diseases were obtained from SONY @ 6300 digital camera and the open source database. Labelme software was used to annotate the diseased spot areas in the images,and after image enhancement, a dataset consisting of 2 250 images was constructed, and divided into training set, validation set and test set according to the ratio of 8 vs 1 vs 1. Three neural network models,Deeplabv3+, U-Net and U-Net++, were trained to identify and segment the images of these three diseases,and the loss curves under the same number of iterations were compared. According to the confusion matrix of predicting the test set, the precision, recall and F1 score of each model were calculated. The results showed that U-Net++ model had the best identification effect, with all three evaluation indicators above97%, and U-Net++ model also had better segmentation performance than Deeplabv3+ and U-Net models.The research results can provide scientific guidance for the disease prevention and control, and also provide technical reference for other crop disease identification.

关 键 词：病害识别 图像 深度学习 模型 预测 U-Net++ 

分 类 号：S432[农业科学—植物病理学] TP391.41[农业科学—农业昆虫与害虫防治] TP18[农业科学—植物保护]