河南农业科学 ›› 2023, Vol. 52 ›› Issue (4): 143-151.DOI: 10.15933/j.cnki.1004-3268.2023.04.017

所属专题: 病虫害识别专题

• 农业信息与工程·农产品加工 • 上一篇    下一篇

基于改进MobileNetV2 模型的农作物叶片病害识别研究

王焕鑫1,沈志豪1,刘泉2,刘金江2   

  1. (1.南阳师范学院生命科学与农业工程学院,河南 南阳 473000;2.南阳师范学院计算机科学与技术学院,河南 南阳 473000)
  • 收稿日期:2022-12-30 出版日期:2023-04-15 发布日期:2023-05-15
  • 通讯作者: 刘金江(1974-),男,河南新乡人,教授,硕士,主要从事图像大数据研究。E-mail:nytc@sina.com
  • 作者简介:王焕鑫(1998-),男,河南周口人,在读硕士研究生,研究方向:图像处理。E-mail:2020086000058@nynu.edu.cn
  • 基金资助:
    河南省教育厅人文社会科学研究项目(2021-ZZJH-262);河南省自然科学基金青年科学基金项目(202300410301)

Identification of Crop Leaf Diseases Based on Improved MobileNetV2 Model

WANG Huanxin1,SHEN Zhihao1,LIU Quan2,LIU Jinjiang2   

  1. (1.College of Life Science and Agricultural Engineering,Nanyang Normal University,Nanyang 473000,China;2.College of Computer Science and Technology,Nanyang Normal University,Nanyang 473000,China)
  • Received:2022-12-30 Published:2023-04-15 Online:2023-05-15

摘要: 为实现基于移动端的农作物叶部病害图像便捷识别,提高农作物病害识别效率进而更好地指导作物病害防治,基于改进的轻量级卷积神经网络MobileNetV2建立农作物病害识别模型。首先,建立含有15 种病害叶片和4 种健康叶片的农作物数据集,采用数据增强操作进行数据平衡。其次,对MobileNetV2 进行改进,引入高效通道注意力(Efficient channel attention,ECA)与注意力特征融合(Attentional feature fusion,AFF),并通过模型剪枝去除冗余层,建立了高性能的轻量级农作物病害识别模型。结果表明:改进MobileNetV2模型参数量与初始MobileNetV2参数量相比减少15.37%,同时识别准确率提升0.9个百分点,达到了98.4%。相比EfficientNet-b0、ShuffleNetV2-0.5X等经典卷积神经网络模型,改进的模型不仅识别准确率最高,且训练过程收敛速度更快。

关键词: MobileNetV2, 卷积神经网络, 农作物病害, 轻量型, 注意力机制, 特征融合, 模型剪枝

Abstract: In order to achieve convenient recognition of crop leaf disease images based on mobile,improve the efficiency of crop disease recognition and thus better guide crop disease control,a crop disease recognition model was established based on an improved lightweight convolutional neural network MobileNetV2.First,a crop dataset containing 15 types of diseased leaves and 4 types of healthy leaves was created and the data was balanced using data enhancement operations.Second,based on the improvement of MobileNetV2,efficient channel attention(ECA)and attentional feature fusion(AFF)were introduced,and the redundant layers were removed by model pruning,and a high⁃performance lightweight crop disease recognition model was proposed.The results showed that the number of parameters of the improved MobileNetV2 model was reduced by 15.37% compared with that of MobileNetV2,while the recognition accuracy was improved by 0.9 percentage points to 98.4% compared with that before the improvement.Compared with the classical convolutional neural network models such as EfficientNet⁃b0 and ShuffleNetV2⁃0. 5X,the improved model not only had the highest recognition accuracy,but also had a faster convergence rate during training.

Key words: MobileNetV2, Convolutional neural network, Crop disease, Lightweight, Attention mechanism, Feature fusion, Model pruning

中图分类号: