Res-Efficient:应用于作物病害自动识别的轻量级模型  

Res-Efficient:A Lightweight Model for Automatic Identification of Crop Diseases

作　　者：管洪亮 李克峰[1] 张广渊[1] 朱振方[1] 王朋[1] GUAN Hong-liang;LI Ke-feng;ZHANG Guang-yuan*;ZHU Zhen-fang;WANG Peng(School of Information Science and Electrical Engineering/Shandong Jiaotong University,Jinan 250375,China)

机构地区：[1]山东交通学院信息科学与电气工程学院,山东济南250375

出　　处：《山东农业大学学报（自然科学版）》2022年第6期858-862,共5页Journal of Shandong Agricultural University：Natural Science Edition

摘　　要：卷积神经网络模型可通过作物病害图像准确率较高地识别作物病害类型,达到防治作物病害的目的,但传统卷积神经网络模型存在模型尺寸大、迁移效果差等问题。针对这些问题,引入学习率动态衰减训练策略,使用EfficientNetV2的Fused-MBConv和MBConv模块替换ResNet18的部分残差模块,提出Res-Efficient模型。实验证明,使用学习率动态衰减策略能提高Res-Efficient模型识别作物病害的准确率,Res-Efficient模型在Plant Village和2018 AI Challenger测试集上分别达到99.70%和87.20%的准确率,模型尺寸减少到14.0 MB。Res-Efficient模型能为移动端和嵌入式设备部署作物病害自动识别应用提供参考。The convolutional neural network model can identify crop disease types with high accuracy through crop disease images to achieve the purpose of preventing and controlling crop diseases.However,the traditional convolutional neural network model has problems such as large model size and poor migration effect.To deal with these problems,a learning rate dynamic attenuation training strategy is introduced,and the Fused-MBConv and MBConv modules of EfficientNetV2 are used to replace part of the residual module of ResNet18,and the Res-Efficient model is proposed.Experiments have proved that using the learning rate dynamic decay strategy can improve the accuracy of the Res-Efficient model in identifying crop diseases.The Res-Efficient model achieved 99.70%and 87.20%accuracy on the Plant Village and 2018 AI Challenger test sets,respectively,and the model size was reduced.to 14.0 MB.The Res-Efficient model can provide a reference for deploying crop disease automatic identification applications on mobile terminals and embedded devices.

关 键 词：作物病害 自动识别 模型 

分 类 号：S41-30[农业科学—植物保护]