农业机械学报2024，Vol.55Issue(3)：252-262,11.DOI:10.6041/j.issn.1000-1298.2024.03.025
基于双节点-双边图神经网络的茶叶病害分类方法
Tea Disease Classification Method Based on Graph Neural Network with Dual Nodes-Dual Edges
摘要
Abstract
The classification of traditional tea diseases mainly relies on manual categorization.Such methods are labor-intensive and time-consuming.Furthermore,insufficient availability of tea disease samples hampers the adequate training of existing machine learning models,resulting in decreased accuracy in disease classification.To address this problem,a tea disease classification method was proposed for four types of tea diseases,including tea anthracnose,tea black rot,and others.This method was based on a dual node-dual edge graph neural network.Firstly,RGB tea disease features and grayscale tea disease features were extracted by using two branches of convolutional neural networks,both branches employed ResNet12 as the backbone network,with independent parameters.The two types of features acted as two sub-nodes within the graph neural network,aiming to obtain disease information from different domains.Secondly,two types of edges,including relative metric edges and similarity edges,were created to improve the aggregation capability of disease features from neighboring nodes.Finally,with the dual node and dual edge feature updating modules,a dual-node and dual-edge alternate updating process was achieved.This process aimed to enhance the accuracy of edge features in measuring node distances.This resulted in achieving accurate classification of tea diseases,even when training samples were limited.Comparative experiments were conducted between the proposed methods,which were based on small-sample learning method.The results indicated that the proposed method achieved superior accuracy in tea disease classification.Specifically,on the miniImageNet and PlantVillage datasets,the proposed method achieved the accuracy of 69.30％and 88.42％in the 5way-1shot,respectively.In the 5way-5shot,the accuracy was improved to 82.48％and 93.04％on the miniImageNet and PlantVillage datasets.Furthermore,on the TeaD-5 tea dataset,the accuracy of the proposed method reached 84.74％in the 5way-1 shot and 86.34％in the 5way-5shot.关键词
茶叶/病害分类/图神经网络/双节点/相对度量边/相似性边Key words
tea/disease classification/graph neural network/dual nodes/relative metric edge/similarity edge分类
信息技术与安全科学引用本文复制引用
张艳,车迅,汪芃,汪玉凤,胡根生..基于双节点-双边图神经网络的茶叶病害分类方法[J].农业机械学报,2024,55(3):252-262,11.基金项目
安徽省中央引导地方科技发展专项(202107d06020001)和国家自然科学基金项目(32372632) （202107d06020001）
