基于改进残差网络的柑橘病害识别  

Citrus disease recognition based on improved residual network

作　　者：帖军[1] 罗均 郑禄[1] 莫海芳[1] 隆娟娟 TIE Jun;LUO Jun;ZHENG Lu;MO Haifang;LONG Juanjuan(College of Computer Science & Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprises,South-Central University for Nationalities,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院&湖北省制造企业智能管理工程技术研究中心,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2021年第6期621-630,共10页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：湖北省技术创新专项重大项目(2019ABA101);武汉市科技计划应用基础前沿项目(2020020601012267);中国科学院-国家民委农业信息技术研究与开发联合实验室招标课题(PJW060012003)。

摘　　要：针对复杂自然环境下柑橘病害图像识别准确率不高的问题,提出一种基于ResNet34深度学习模型的多类别柑橘病害图像识别方法.通过舍弃部分残差结构中的identity映射,进一步提取柑橘病害图像的低层特征,降低消极特征在低层特征的占比,得到改进后的S-ResNet模型.与ResNet34模型相比:S-ResNet模型对自然环境下柑橘病害图像的识别准确率提高了3.9%.为提取柑橘病害图像中更具表达力的深层特征,使用3×3卷积核替换ResNet34首层中的7×7卷积核,得到改进后的M-ResNet模型.使用模型融合方法将S-ResNet与M-ResNet进行融合,得到融合模型F-ResNet,解决了单一模型中泛化能力弱、鲁棒性较差等问题.实验结果表明,F-ResNet对自然环境下柑橘病害图像的识别准确率达到93.6%,可以很好地满足实际生产环境中果园病害识别需求,具有良好的应用前景.In order to solve the problem that the recognition accuracy of citrus disease is not high in complex natural environment,we present a multi-classification citrus disease image recognition method based on ResNet34 deep learning model.An improved model S-ResNet was proposed by discarding some identity mappings in residual structure of ResNet34.In S-ResNet,the low layers features of the citrus disease images are further extracted,and the negative features are reduced.Compared with ResNet34 model,the average recognition accuracy of S-ResNet model is improved by 3.9%.In addition,we propose a model called M-ResNet using three 3×3 convolution kernels instead of the original 7×7 convolution kernels was introduced to extract more expressive deep features of citrus disease images.To avoid the problems of bad training approximation,generalization and robustness in single model,S-ResNet and M-ResNet are fused by model fusion method to obtain a model called F-ResNet.The experimental results show that the accuracy rate of F-ResNet can reach 93.6%,and the recognition accuracy of citrus disease image in natural environment can better meet the needs of disease identification in actual orchard,and has a good application prospect.

关 键 词：病害识别 ResNet34模型 融合模型 卷积神经网络 

分 类 号：S24[农业科学—农业电气化与自动化] TP2[农业科学—农业工程]