English

智能化农业装备学报(中英文) ›› 2021, Vol. 2 ›› Issue (1): 64-70.DOI: 10.12398/j.issn.2096-7217.2021.01.009

• • 上一篇    

基于卷积神经网络的玉米病害识别方法研究*

王国伟, 刘嘉欣   

  1. 吉林农业大学信息技术学院,长春市,130000
  • 收稿日期:2020-05-13 出版日期:2021-05-15 发布日期:2021-11-30
  • 作者简介:王国伟,男,1977年生,吉林长春人,硕士,副教授;研究方向为人工智能、精准农业。E-mail: 41422306@qq.com
  • 基金资助:
    *2017年度吉林省科技发展计划项目(20170204020NY)

Research on corn disease recognition method based on convolutional neural network*

Guowei Wang, Jiaxin Liu   

  1. College of Information Technology, Jilin Agricultural University, Changchun, 130000, China
  • Received:2020-05-13 Online:2021-05-15 Published:2021-11-30
  • About author:Guowei Wang, Associate Professor, research direction: artificial intelligence and precision agriculture. E-mail: 41422306@qq.com
  • Supported by:
    *2017 Jilin Science and Technology Development Plan Project (20170204020NY)

摘要: 为解决传统的玉米病害识别方法中特征提取主观性强及误识率高的问题,提出利用卷积神经网络对玉米病害进行识别。以玉米病害图像和健康图像共5种类别的玉米图像为研究对象,并采用LeNet模型进行试验。首先,按照8∶2的比例为每种玉米病害图像选择训练集和测试集。然后,通过试验组合和对比分析的方法比较不同卷积神经网络结构设置对准确率的影响,选出最佳参数。另外,选用Adam算法代替SGD算法来优化模型,通过指数衰减法调整学习率,将L2正则项添加到交叉熵函数中,并选择Dropout策略和ReLU激励函数。最后,确定了一个10层CNN网络结构。试验结果显示,玉米花叶病、灰斑病、锈病、叶斑病和玉米健康识别率分别为95.83%、90.57%、100%、93.75%、100%,平均识别率达96%,平均计算时间为0.15 s。经试验结果比较,该模型识别效果明显高于传统方法,为玉米病害的防治提供技术支持。

关键词: 卷积神经网络, 玉米图像, LeNet模型, 病害识别

Abstract: In order to solve the problems of strong subjective feature extraction and high misrecognition rate in traditional corn disease identification methods, a convolutional neural network is proposed to identify corn diseases. The corn images with 5 types of corn disease images and healthy images were taken as the research objects, and the LeNet model was used for experiments. First, select a training set and a test set for each corn disease image in the ratio of 8∶2. Then, the effects of different convolutional neural network structure settings on accuracy are compared through experimental combination and comparative analysis, and the best parameters are selected. In addition, the Adam algorithm is used instead of the SGD algorithm to optimize the model, the learning rate is adjusted by the exponential decay method, the L2 regular term is added to the cross entropy function, and the Dropout strategy and ReLU excitation function are selected. Finally, a 10-layer CNN network structure is determined. The experimental results show that the maize mosaic disease, gray leaf spot, rust, leaf spot disease and corn health recognition rates are 95.83%, 90.57%, 100%, 93.75%, 100%, and the average recognition rate is 96%. Compared with experimental results, the recognition effect of this model is significantly higher than traditional methods, which provides technical support for the prevention and treatment of corn diseases.

Key words: convolutional neural network, corn image, LeNet model, disease recognition

中图分类号: