基于卷积神经网络的农作物病虫害检测研究  

Research on Crop Diseases and Pests Detection Based on Convolutional Neural Network

作　　者：白雪松 吴建平[1,2,3] 景文超[1] 何旭鑫 余咏 BAI Xue-song;WU Jian-ping;JING Wen-chao;HE Xu-xin;YU Yong(School of Information Science&Engineering,Yunnan University,Kunming 650504,China;Yunnan Provincial Electronic Computing Center,Kunming 650223,China;Key Laboratory of Digital Media Technology of Universities and Colleges in Yunnan Province,Kunming 650223,China)

机构地区：[1]云南大学信息学院,云南昆明650504 [2]云南省电子计算中心,云南昆明650223 [3]云南省高校数字媒体技术重点实验室,云南昆明650223

出　　处：《计算机技术与发展》2022年第12期200-205,共6页Computer Technology and Development

基　　金：云南省重大科技专项计划项目(202002AD080001);云南省科技厅应用基础研究计划重点项目(2019FA044);云南大学第一届专业学位研究生实践创新项目(2021Y183)。

摘　　要：农作物病虫害图像采集困难,且公共数据集较少,因此使用单一网络模型识别准确率不高。常用的数据增强方法只能对图像进行像素空间的变换,不能进行语义转换。对此,研究并提出基于隐式语义数据增强算法的CBAM-Res2Net50模型:该模型使用Res2Net50主干网络,从多尺度学习图像信息,加载预训练模型部分参数,提高模型的收敛速度;在网络残差块中添加混合注意力模块,提取并保留关键特征;训练过程中使用隐式语义数据增强算法对提取的深层网络空间特征进行语义扩充增强,提高模型的泛化能力。改进模型与现有模型在AI Challenger 2018农作物病虫害数据集上的对比实验结果表明:改进模型具有较高的识别率,其分类准确率达88.33%。改进后的模型通过挖掘相似病虫害图像的语义信息,在一定程度上解决了深度网络中由于数据不足导致的过拟合等问题。It is difficult to collect images of crop diseases and pests,and there are few common data sets,so the recognition accuracy of single network model is not high.The commonly used data augmentation methods can only transform image pixel space,but not semantic transformation.Therefore,CBAM-Res2Net50 model based on implicit semantic data augmentation algorithm is studied and proposed.The model uses Res2Net50 backbone network to learn image information from multi-scale and load some parameters of pre-training model to improve the convergence speed of the model.A mixed attention module is added to the network residual block to extract and retain key features.In the training process,implicit semantic data augmentation algorithm is used to enhance semantic expansion and generalization ability of the model.The experimental results of comparison between the improved model and the existing model on AI Challenger 2018 Crop Diseases and Pests Data Set show that the improved model has a high recognition rate,and its classification accuracy is 88.33%.The improved model can solve the problem of over-fitting caused by insufficient data in deep network to some extent by mining semantic information of similar images of diseases and pests.

关 键 词：农作物病虫害 卷积神经网络 CBAM-Res2Net50 迁移学习 注意力模块 隐式语义数据增强 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]