English

中国农机化学报

中国农机化学报 ›› 2023, Vol. 44 ›› Issue (8): 191-197.DOI: 10.13733/j.jcam.issn.2095-5553.2023.08.026

• 农业智能化研究 • 上一篇    下一篇

基于迁移学习和卷积视觉转换器的农作物病害识别研究

余胜,谢莉   

  1. 韶关学院信息工程学院,广东韶关,512005
  • 出版日期:2023-08-15 发布日期:2023-09-12
  • 基金资助:
    广东省自然科学基金项目(2021A1515011803);韶关市科技计划项目(210728114530796、210728104530586)

Research on plant disease identification based on transfer learning and convolutional vision transformer

Yu Sheng, Xie Li   

  • Online:2023-08-15 Published:2023-09-12

摘要: 农作物病虫害对粮食生产和质量都有很大影响。针对当前传统的农作物病害识别过程中主要依赖人工特征提取,且真实环境下采集的病害图像样本数目较少,识别方法鲁棒性差、分类准确率偏低等问题,基于迁移学习提出了以卷积操作预处理图像子块的视觉转换器(Vision Transformer,ViT)模型用于农作物病害识别。在ViT模型结构的基础上引入卷积操作对输入图像进行预处理,卷积操作能提高获取底层特征的丰富度,进而在ViT学习过程中通过多头注意力机制,加大有用特征的权重,削弱噪声等无用信息的影响,达到模型提高特征学习能力并增强鲁棒性的目的。试验结果表明,利用迁移学习方法在ibean数据集上能够提升模型的识别准确率10%以上;模型最终在ibean数据集上识别准确率为98.12%,约有2%的提高,在PlantVillage数据集识别准确率为99.91%,都达到了当前最佳识别水平。提出的识别方法在复杂背景干扰下具有较高的识别准确率和鲁棒性,可以满足自然条件下的农作物病害识别的要求。

关键词: 迁移学习, 卷积视觉转换器, 病害识别, 注意力机制

Abstract: The plant diseases impact on both the food production and quality in the agriculture sector. As the problem that traditional plant disease identification methods rely on training sample data and manual extraction of features, it is difficult to identify in the field environment and the classification accuracy is not high. Based on transfer learning, this study proposes a Convolutional Vision Transformer (CViT) model that preprocesses image subblocks with convolution operations for plant disease identification. On the basis of the Vision Transformer (ViT) model structure, the convolution operation is introduced to preprocess input images, the convolution operation improves the richness of the lowlevel features, and then in the ViT learning process, the multihead mechanism is used to increase the weight of useful features and noise suppression. So as to achieve the purpose of improving the feature learning ability and enhancing the robustness of the model. The experimental results show that using the transfer learning method on the ibean dataset can improve the recognition accuracy of the model by more than 10%. Applying transfer learning to the CViT model achieves a recognition accuracy of 98.12% on the ibean dataset, an improvement of about 2%. The recognition accuracy in the PlantVillage dataset is 99.91%. The proposed recognition method has high recognition accuracy and robustness under complex background interference, and can meet the requirements of plant disease identification under natural conditions.

Key words: transfer learning, convolutional vision transformer, plant disease identification, attention mechanism

中图分类号: