English

中国农机化学报

中国农机化学报 ›› 2022, Vol. 43 ›› Issue (11): 188-194.DOI: 10.13733/j.jcam.issn.2095-5553.2022.11.026

• 农业智能化研究 • 上一篇    下一篇

基于卷积神经网络的害虫分类

陈继清,韦德鹏,龙腾,罗天,王桦彬   

  1. 广西大学机械工程学院,南宁市,530007
  • 出版日期:2022-11-15 发布日期:2022-10-25
  • 基金资助:
    广西科技基础和人才专项(AD19110034)

Pest classification based on convolutional neural network

Chen Jiqing, Wei Depeng, Long Teng, Luo Tian, Wang Huabin.   

  • Online:2022-11-15 Published:2022-10-25

摘要: 农作物病虫害是一种严重的自然灾害,需要对其进行及时预测和监控,以保证农作物产量。由于害虫种类繁多以及作物在生长初期的形态相似,农业工作者难以准确识别各类作物昆虫,给病虫害的防治工作带来巨大挑战。针对这一问题,提出一种基于多尺度特征融合的网络模型(FFNet)对作物害虫进行精准识别与分类。首先,采用空洞卷积设计多尺度特征提取模块(MFEM),获取害虫图像的多尺度特征图;然后,使用深层特征提取模块(DFEM)提取图像的深层特征信息;最后,将分别由多尺度特征提取模块(MFEM)和深层特征提取模块(DFEM)提取到的特征图进行融合,从而实现以端到端的方式对作物害虫进行精准分类与识别。试验表明:所提出的方法在12类害虫的数据集上获得优异的分类性能,分类准确率(ACC)达到98.2%,损失函数Loss为0.031,模型训练时间为197 min。

关键词: 深度学习, 卷积神经网络, 多尺度融合, 空洞卷积, 害虫分类

Abstract: Crop diseases and insect pests are serious natural disasters, which need to be predicted and monitored in time to ensure crop yields. Due to the wide variety of pests and the similar shapes of crops in the early stage of growth, it is difficult for agricultural workers to accurately identify various crop insects, which brings great challenges to the prevention and control of pests and diseases. Aiming at this problem, a network model based on multiscale feature fusion (FFNet) was proposed to accurately identify and classify crop pests. First, a multiscale feature extraction module (MFEM) was designed by using cavity convolution to obtain the multiscale feature map of the pest image; then, the deep feature extraction module (DFEM) was used to extract the deep feature information of the image. The feature maps extracted by the extraction module (MFEM) and the deep feature extraction module (DFEM) were fused to achieve accurate classification and identification of crop pests in an endtoend manner. Experiments showed that the proposed method achieved excellent classification performance on the dataset of 12 types of pests, the classification accuracy rate (ACC) reached 98.2%, the loss function Loss was 0.031, and the model training time was 197 min respectively.

Key words: deep learning, convolutional neural network, multiscale fusion, dilated convolution, pest classification

中图分类号: