计算机工程与应用 ›› 2022, Vol. 58 ›› Issue (12): 260-268.DOI: 10.3778/j.issn.1002-8331.2111-0457

• 工程与应用 • 上一篇    下一篇

改进ShuffleNet V2的轻量级农作物病害识别方法

李好,邱卫根,张立臣   

  1. 广东工业大学 计算机学院,广州 510006
  • 出版日期:2022-06-15 发布日期:2022-06-15

Improved ShuffleNet V2 for Lightweight Crop Disease Identification

LI Hao, QIU Weigen, ZHANG Lichen   

  1. School of Computers, Guangdong University of Technology, Guangzhou 510006, China
  • Online:2022-06-15 Published:2022-06-15

摘要: 针对目前有关深度学习的农作物病害识别方法中存在模型较为复杂,部署在计算资源有限的边缘设备和移动终端上适应性不强,实时准确识别作物病害较差的问题,提出一种改进ShuffleNet V2的轻量级农作物病害识别方法。以ShuffleNet V2单元为基础,引入ECA(efficient channel attention)注意力模块,使用H-Swish激活函数以便减少网络结构每个Stage模块中ShuffleNet V2单元使用个数,使用轻量化网络设计组件深度可分离卷积。在PlantVillage病害数据集上进行实验。结果表明,模型的参数量约为2.95×105,计算量为3.388×107(FLOPs)和6.674×107(MAdd),病害识别平均准确率达到了99.24%,为农作物病害识别方法在移动终端等资源受限设备上部署应用提供参考。

关键词: 农作物病害识别, ShuffleNet V2, 轻量级, ECA注意力模块

Abstract: Aiming at the problems of current crop disease deep learning identification methods are more complex, poor adaptability for deployment on edge mobile devices with limited computing resources and poor real-time accurate identification, a lightweight crop disease identification method based on improved ShuffleNet V2 is proposed. Based on ShuffleNet V2 Unit, the efficient channel attention(ECA) attention module is introduced, using  H-Swish activation function inorder to reduce ShuffleNet V2 Unit numbers in each Stage module of network structure, and using a lightweight network design component with depthwise separable convolution. Finally, experiments are conducted on the PlantVillage disease dataset. Experimental results show that model parameter is 2.95×105, computational volume is 3.388×107(FLOPs) and 6.674×107(MAdd), and crop disease identification average accuracy reaches 99.24%, provides a reference for the deployment of crop diseases identification methods on resource-constrained devices such as mobile terminals.

Key words: crop disease identification, ShuffleNet V2, lightweight, efficient channel attention(ECA) attention module