第22卷第2期2023年2月Vol.22No.2Feb.2023软件导刊SoftwareGuide嵌入压缩—激励机制ResNet的民族药植物图像识别周婷,杜建强,朱彦陈,冯振乾(江西中医药大学计算机学院,江西南昌330004)摘要:针对民族药植物图像数据集稀缺、样本量少、图像背景复杂以致图像特征提取困难的问题,构建TibetanMP数据集,提出一种嵌入压缩—激励机制ResNet并结合迁移学习的图像识别方法。该方法对ResNet34在ImageNet上的预训练模型进行迁移学习,以减少过拟合现象,同时在网络浅层中引入SE机制,使网络聚焦图像中的关键特征,最后对模型进行微调。为了评估所提方法的性能,在TibetanMP、Oxford102flowers和CIFAR-10数据集上进行实验,模型分别取得96.33%、98.81%和91.92%的识别准确率。与其他主流CNN图像识别模型进行比较,发现该模型具有更高的识别精度,具有一定的工程实用性。关键词:图像识别;民族药植物图像;ResNet;压缩激励;迁移学习DOI:10.11907/rjdk.221181开放科学(资源服务)标识码(OSID):中图分类号:TP391.4文献标识码:A文章编号:1672-7800(2023)002-0001-07ImageRecognitionofEthnicMedicinalPlantsEmbeddingSqueeze-and-ExcitationMechanismResNetZHOUTing,DUJian-qiang,ZHUYan-chen,FENGZhen-qian(CollegeofComputerScience,JiangxiUniversityofChineseMedicine,Nanchang330004,China)Abstract:Aimingattheproblemsofscarcityofethnicmedicinalplantsimagedataset,smallsamplesizecompleximagebackground,whichmakeimagefeatureextractiondifficult,theTibetanMPdatasetisconstructedandanimagerecognitionmethodwithembeddedSqueeze-and-ExcitationmechanismResNetcombinedwithtransferlearningwasproposed.Inthismethod,thepre-trainingmodelofResNet34onImageNetistransferredtoreducetheover-fittingphenomenon.Meanwhile,SEmechanismisintroducedintheshallowlayerofthenetworktofocusthekeyfeaturesintheimage.Finally,themodelisfine-tuned.Inordertoevaluatetheperformanceoftheproposedmethod,ontheTibetanMP,Oxford102flowersandCIFAR-10datasets,themodelachievedrecognitionaccuracyof96.33%,98.81%and91.92%,respectively.Com⁃paredwithothermainstreamCNNimagerecognitionmodels,thismodelhashigherrecognitionaccuracy.Experimentsshowthatthismethodcaneffectivelyimprovetheimagerecognitionpe...