English

中国农机化学报

中国农机化学报 ›› 2024, Vol. 45 ›› Issue (2): 250-258.DOI: 10.13733/j.jcam.issn.2095-5553.2024.02.036

• 农业智能化研究 • 上一篇    下一篇

基于分层特征对齐网络的小样本马铃薯病害叶片检测

牛玉霞1, 2,孙宙红3,任伟1, 2,陈林琳1, 2,陈莉莉1, 2   

  • 出版日期:2024-02-15 发布日期:2024-03-20
  • 基金资助:
    江苏省农业农村污染防治技术与装备工程研究中心开发资金资助(GCZXYB2305)

Few-shot potato disease leaf detection based on hierarchical feature alignment network

Niu Yuxia1, 2, Sun Zhouhong3, Ren Wei1, 2, Chen Linlin1, 2, Chen Lili1, 2   

  • Online:2024-02-15 Published:2024-03-20

摘要: 针对传统马铃薯病害叶片检测方法过度依赖大量训练数据以及对未知病害识别泛化性不强的问题,提出一种基于分层特征对齐网络的小样本马铃薯病害叶片检测模型。首先,收集并整理包含多种病害类型的弱标注马铃薯病害叶片数据集。其次,在支持分支中建立文本语义和视觉语义的多模态双层特征语义表示,并利用预训练网络生成多个候选框。再次,利用卷积神经网络将候选框区域映射到深度特征空间,并借助无参数的度量方法实现文本语义与视觉语义的特征对齐。最后,将查询分支中的未知类病害图片与多模态视觉和文本语义关联集进行度量计算,根据相似度值快速给出待测图片中未知新类的病害类别。通过在自建的马铃薯病害叶片数据集和开源数据集上进行测试,所提出模型分别可以实现93.55%和96.35%的识别精度,在跨域数据集上可以实现95.15%和94.06%的识别精度,优于当前经典的目标检测模型,具有一定的实际应用价值。

关键词: 马铃薯病害, 叶片检测, 分层特征对齐网络, 文本语义, 视觉语义

Abstract: In order to address the problems of the over-reliance on large amounts of training data and the poor generalization of unseen disease identification in traditional potato disease leaf detection methods, a few-shot potato disease leaf detection model based on hierarchical feature alignment network is proposed. Firstly, a weakly labeled dataset containing various types of potato diseases were collected and annotated. Secondly, the multi-modal bi-modal feature semantic representations of textual and visual semantics in the support branch were established, and multiple candidate boxes were generated using a pre-trained region proposal network. Thirdly, a convolutional neural network was adopted to map the candidate box regions into deep feature space, and feature alignment was performed between textual and visual semantics using an unparameterized metric method. Finally, the similarity was computed between the unseen class disease images in the query branch and the multi-modal visual and textual semantic association set, and the disease category of the unseen new class was quickly provided according to the similarity value. Through testing on self-built potato disease leaf datasets and open source datasets, the proposed models can achieve recognition accuracy of 93.55% and 96.35% on the test sets, respectively, and 95.15% and 94.06% on the cross-domain datasets, which is superior to the current classical object detection models. The proposed method has certain practical application value.

Key words: potato disease, leaf detection, hierarchical feature alignment network, textual semantics, visual semantics

中图分类号: