基于改进MobileNetV2的人脸表情识别*严春满,张翔,王青朋(西北师范大学物理与电子工程学院,甘肃兰州730070)摘要:针对现有深度卷积神经网络参数量庞大,导致人脸表情识别场景受限的问题,提出一种基于改进轻量级卷积神经网络的人脸表情识别模型。该模型以MobileNetV2轻量级特征提取网络为主要框架,通过压缩网络宽度因子与整体维度,减少网络参数量与计算量;引入SandGlass模块对网络倒残差模块进行改进,减少特征信息在网络传输中的丢失;同时嵌入高效通道注意力机制,提高网络对于特征信息的提取能力。在人脸表情数据集FER2013和CK+上进行实验,所提网络模型的人脸表情识别准确率达到了68.96%与95.96%,分别高于MobileNetV21.06%与6.14%,且参数量下降82.28%,实验结果验证了网络模型改进措施的有效性。关键词:人脸表情识别;轻量级网络;MobileNetV2;倒残差模块;通道注意力中图分类号:TP391.41文献标志码:Adoi:10.3969/j.issn.1007-130X.2023.06.014FacialexpressionrecognitionbasedonimprovedMobileNetV2YANChun-man,ZHANGXiang,WANGQing-peng(CollegeofPhysicsandElectronicEngineering,NorthwestNormalUniversity,Lanzhou730070,China)Abstract:Aimingattheproblemthattheexistingdeepconvolutionalneuralnetworkhasalargea-mountofparameters,whichleadstothelimitationoffacialexpressionrecognitionscenes,thispaperproposesafacialexpressionrecognitionmodelbasedonimprovedlightweightconvolutionalneuralnet-work.ThemodeltakesMobileNetV2lightweightfeatureextractionnetworkasthemainframework,bycompressingthenetworkwidthfactorandtheglobaldimension,thenumberofnetworkparametersandtheamountofcomputationarereduced.SandGlassblockisintroducedtoimprovethereverseresidualmoduleinthisnetwork,andreducethelossoffeatureinformationduringnetworktransmission.Atthesametime,theefficientchannelattentionmechanismisembeddedtoimprovethenetwork'sabilitytoex-tractfeatureinformation.ExperimentswerecarriedoutonthefacialexpressiondatasetsFER2013andCK+.Thefacialexpressionaccuracyrateoftheproposednetworkreaches68.96%and95.96%,whichare1.06%and6.14%higherthanthatofMobileNetV2respectively,andthenumberofparametersaredecreasedby82.28%.Experimentalresultsverifytheeffectivenessoftheimprovednetworkmodel.Keywords:fac...
