基于深度学习的水稻病虫害诊断方法研究  

Research on Rice Diseases and Pests Diagnosis Based on Deep Learning

作　　者：姜敏 沈一鸣 张敬尧 饶元[1] 董伟 JIANG Min;SHEN Yiming;ZHANG Jingyao;RAO Yuan;DONG Wei(Anhui Agricultural University,Hefei 230036,China;Anhui Academy of Agricultural Sciences,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]安徽省农业科学院农业经济与信息研究所,安徽合肥230036

出　　处：《洛阳理工学院学报（自然科学版）》2019年第4期78-83,共6页Journal of Luoyang Institute of Science and Technology：Natural Science Edition

基　　金：国家自然科学基金资助项目(61572260)

摘　　要：病虫害是影响水稻等农作物产量的重要制约性因素。为探索基于深度学习的水稻病虫害诊断方法,采用图片尺寸归一化、截取感兴趣区域、病理分割3种预处理方式分别与Faster R-CNN Inception v2、SSD MobileNet v1两种深度学习目标检测预训练模型结合,在TensorFlow深度学习平台下进行水稻病虫害识别模型的训练和诊断效果测试。实验结果表明,6种条件下水稻病虫害识别准确率分别为99.65%、90.74%、92.60%、82.23%、65.74%和20.41%,其中采用归一化尺寸和Faster R-CNN模型时水稻病虫害识别准确率最高,且具有较低的训练时长,较适宜用于水稻病虫害诊断。Diseases and pest damage are the important factors of yield of rice and other crops. In order to investigate deep-learning-based diagnosis method of rice pests and diseases, the different combinations of three image preprocessing methods and two deep-learning-based target detection algorithms are implemented. That is normalizing image size, truncating region of interest and pathological segmentation combined with Faster R-CNN Inception v2 or SSD MobileNet v1 respectively. The identification models of rice diseases and pests are trained and tested with Tensorflow platform. The experimental results show that the identification rate of rice diseases and pests for six combinations are 99.65%, 90.74%, 92.60%, 82.23%,65.74% and 20.41%,respectively, among which the combination of normalizing image size and Faster R-CNN model has the highest identification precision and rate, and lower training time. This implies that the aforementioned combination is more suitable for conducting the diagnosis of rice diseases and pests compared with other combinations.

关 键 词：深度学习 水稻 病虫害识别 诊断 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]