第27卷第6期2022年12月哈尔滨理工大学学报JOURNALOFHARBINUNIVERSITYOFSCIENCEANDTECHNOLOGYVol.27No.6Dec.2022融合并行网络特征的人脸表情识别算法苗壮1,程卫月2,林克正1,李骜1(1.哈尔滨理工大学计算机科学与技术学院,哈尔滨150080;2.黑龙江工商学院,哈尔滨150025)摘要:针对单一卷积神经网络对人脸表情特征提取不充分和参数量较大等问题,提出了一种融合并行网络特征的人脸表情识别算法。该算法首先对ResNet网络中的残差块进行修改,减少网络参数量同时使用预激活来减小错误率。之后将改进后的ResNet网络提取到的特征与剪层后的VGG网络提取到的特征进行融合,得到网络模型P-ResNet-VGG,其中损失函数使用交叉熵损失函数。该模型已在FER2013和JAFFE数据集上进行了大量实验。实验结果表明,该模型比其他几种模型在FER2013和JAFFE表情数据集上准确率都有所提高,具有更好的鲁棒性。关键词:深度学习;卷积神经网络;人脸表情识别;并行网络;特征融合DOI:10.15938/j.jhust.2022.06.012中图分类号:TP391.4文献标志码:A文章编号:1007-2683(2022)06-0095-08FacialExpressionRecognitionAlgorithmCombiningParallelNetworkFeaturesMIAOZhuang1,CHENGWei-yue2,LINKe-zheng1,LIAo1(1.SchoolofComputerScienceandTechnology,HarbinUniversityofScienceandTechnology,Harbin150080,China;2.HeilongjiangCollegeofBusinessandTechnology,Harbin150025,China)Abstract:Aimingattheproblemsofinsufficientextractionoffacialexpressionfeaturesbyasingleconvolutionalneuralnetworkandlargeamountofparameters,afacialexpressionrecognitionalgorithmfusedwithparallelnetworkfeaturesisproposed.ThealgorithmfirstmodifiestheresidualblockintheResNetnetwork,reducestheamountofnetworkparametersandusespre-activationtoreducetheerrorrate.Afterthat,thefeaturesextractedbytheimprovedResNetnetworkandthefeaturesextractedbytheVGGnetworkafterthecutlayeraremergedtoobtainthenetworkmodelP-ResNet-VGG,inwhichthelossfunctionusesthecross-entropylossfunction.ThismodelhasbeenextensivelytestedontheFER2013andJAFFEdatasets.ExperimentalresultsshowthatthismodelhasimprovedaccuracyonFER2013andJAFFEexpressiondatasetsthanothermodels,andhasbetterrobustness.Keywords:deeplearning;convolutionalneuralnetwo...