多分支特征级联小样本农作物病害图像识别
DOI:
作者:
作者单位:

重庆师范大学 计算机与信息科学学院

作者简介:

通讯作者:

基金项目:

重庆市高校创新研究群体项目(CXQT20015)


Multi-branch feature cascaded few-shot crop diseases image recognition
Author:
Affiliation:

School of Computer and Information Science, Chongqing Normal University

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
    摘要:

    【目的】针对现有小样本学习农作物病害识别参数量高、网络易出现梯度消失等问题,提出一种多分支分组卷积的特征级联小样本学习农作物病害图像识别模型。【方法】首先,该模型将分组卷积应用于残差块,利用残差块缓解网络的梯度消失问题。其次,将网络的输入拆分成同构多分支的嵌入,通过不同分支间的特征补充,获取更加丰富的特征并同时降低网络参数量。最后,设计了一种多层特征级联的骨干网络,用于挖掘图像中不同层次的信息,将浅层特征中更多位置、细节信息和深层特征中更强的语义信息级联输出,提升网络的泛化能力。【结果】在PlantVillage数据集在5-way 1-shot与5-way 5shot的准确率分别为79.68±0.35、93.25±0.21,比原型网络高出3.74%和4.32%。【结论】本文模型在农作物病害识别的有效性有一定的性能提升,并且实现了加深网络深度的同时尽可能降低网络参数量、缓解梯度消失、挖掘不同层次的语义信息等作用,能有效解决农作物病害有效样本较少的问题。

    Abstract:

    [Purposes] To address the problems of high parameter quantities and gradient disappearance in existing few-shot learning for summarizing crop disease recognition, a multi-branch grouped convolutional feature cascade few-shot learning model for crop disease image recognition is proposed. [Methods] Firstly, the model applies grouped convolution to the residual block, using the residual block to alleviate the gradient disappearance problem of the network. Secondly, the input of the network is split into isomorphic multi-branch embeddings, and the feature supplement between different branches is used to obtain more rich features and reduce the number of network parameters at the same time. Finally, a multi-layer feature cascade backbone network is designed to mine information at different levels in the image, cascading the output of more location, detail information in shallow features and stronger semantic information in deep features to improve the generalization ability of the network. [Finding] In the PlantVillage dataset, the accuracy rates of 5-way 1-shot and 5-way 5shot are 79.68±0.35 and 93.25±0.21, respectively, which are 3.74% and 4.32% higher than the prototype network. [Conclusions] The model in this paper has a certain performance improvement in the effectiveness of crop disease identification, and realizes the functions of deepening the network depth while reducing the amount of network parameters as much as possible, alleviating gradient disappearance, and mining semantic information at different levels, which can effectively solve the problem of effective samples of crop diseases. fewer problems.

    参考文献
    相似文献
    引证文献
引用本文
分享
文章指标
  • 点击次数:
  • 下载次数:
历史
  • 收稿日期:2023-02-24
  • 最后修改日期:2023-09-21
  • 录用日期:2024-09-04
  • 在线发布日期: