第46卷第1期2023年2月电子器件ChineseJournalofElectronDevicesVol.46No.1Feb.2023项目来源:国家自然科学基金资助项目(62076215);江苏省高等学校自然科学研究重大项目资助(19KJA110002)收稿日期:2022-04-03修改日期:2022-06-13IdentificationofDrugAddictsbyDeepLearningandSVM*WANGYuanyuan1*,XUYide2,WANGXinyu1,TIANBin3,WANGKuiwen3,ZHOUFeng1(1.SchoolofInformationTechnology,YanchengInstituteofTechnology,YanchengJiangsu224051,China;2.SchoolofInformationScienceandEngineering,SoutheastUniversity,NanjingJiangsu210096,China;3.JiangsuFangqiangCompulsoryIsolatedDetoxificationCenter,YanchengJiangsu224165,China)Abstract:AneuralnetworkmodelusingPCAandlineardiscriminatorisproposedtoidentifyaddictionlevelandcommunitycorrectiontimeofdrugaddictsaccordingtofacialimagesaswellaslocalinformationinabstractfeaturesthroughdeeplearningmethod.Firstly,thebackbonenetworkResNet50ispre-trained.Then,PCAisusedtoreducethenumberoffeaturesandtheFisherdiscriminatorisusedforpre-discrimination,sothat,thetrainingtimeofthemodelisreducedandfeatureextractionismoreaccurateandfaster.Finally,thenet-workendisclassifiedthroughthecombinationofthefullyconnectedlayerandtheSVMfunction.Cross-entropylossisadoptedastheoptimizationgoalofstochasticgradientdescent.Theexperimentalresultsshowthatthemethodhasarecognitionaccuracyof81.74%forthedegreeofdrugaddictionand60.59%forthecommunitycorrectiontime.Keywords:deepLearning;neuralnetworks;PCA;fisher’slineardiscriminantEEACC:6135doi:10.3969/j.issn.1005-9490.2023.01.020基于深度学习与SVM的吸毒成瘾者识别*王媛媛1*,徐一得2,王新宇1,田彬3,王奎文3,周锋1(1.盐城工学院信息工程学院,江苏盐城224051;2.东南大学信息科学与工程学院,江苏南京210096;3.江苏省方强强制隔离戒毒所,江苏盐城224165)摘要:提出一种使用PCA和线性判别器的神经网络模型,利用深度学习方法通过面部图像及抽象特征中的局部信息识别吸毒成瘾者的成瘾程度和社区矫正时间。首先对主干网络ResNet50进行预训练;再使用PCA降低特征数、Fisher判别器进行预判,从而使模型的训练时间减少、特征提取更加准确和快捷;最后网络末端...