基于深度卷积神经网络的玉米病害识别.PDFVIP

中国农学通报 2018,34(36): 159-164 Chinese Agricultural Science Bulletin 基于深度卷积神经网络的玉米病害识别 刘永波，雷 波，曹 艳，唐江云，胡 亮 （四川省农业科学院农业信息与农村经济研究所，成都610011） 摘 要：为了提高玉米病害的识别率，本文提出了一种在自然环境条件下基于深度卷积神经网络的玉米 病害识别方法。该方法以玉米常见的10类病害为研究对象。算法模型是先将图像预处理，应用Triplet loss 双卷积神经网络结构学习玉米图像特征，再使用SIFT 算法提取图像纹理细节，最后通过Softmax 对 图像进行标签分类。训练集采用正常玉米图像与玉米病害图像相结合的方式，使用深度相似性网络学 习正常玉米图像特征表示，再使用迁移学习方法学习玉米病害图像的特征，最后对特征进行分类识别。 研究结果表明：该方法可准确识别10种常见玉米病害，正确率可达90% 以上，为玉米病害的防治提供了 有效的技术支持。 关键词：玉米病害；算法；识别；深度学习；卷积神经网络；图像处理 中图分类号：TP391 文献标志码：A 论文编号：cas MaizeDiseasesIdentificationBasedonDeepConvolutionalNeuralNetwork LiuYongbo,LeiBo,CaoYan,TangJiangyun,HuLiang (InstituteofAgriculturalInformationandRuralEconomy,SichuanAcademyofAgriculturalSciences,Chengdu610011) Abstract: To improve the identification rate of maize diseases, an identification method of diseases is establishedbasedonadeepconvolutionalneuralnetwork under natural environment condition.Tencommon maizediseaseswere takenasthestudyobjects,andtheimagepreprocessingwascarriedout,theTripletloss doubleconvolutionneuralnetworkstructurewasapplied tostudy thefeaturesofmaizeimages.Then,SIFT algorithmwasadoptedtoextracttexturalfeatures.Bythisway,thelabelingandclassificationofimagesthrough theSoftmaxwere conducted.Thenormalmaizeimagesandthediseasedoneswerecombinedbythetraining set.Deeplearningonsimilaritynetworkcheckwasusedtoexplorethefeaturesofnormalmaizeimages,and furthermore,transferlearningwasusedtodelveintothefeaturesofdiseasedmaizeimages.Thefeatureswere classifiedandidentified.Theresultshowsthatthismethodcouldaccuratelyidentifythetencommonmaize diseases,andtheclassificationaccuracyisover90%. Key words: maizedisease;algorithms;identification;deeplearning;convolutionalneuralnetwork;image processing; 0 引言