基于DenseNet的人脸图像情绪识别研究  

Research on emotion recognition of face image based on densenet

作　　者：雷建云[1] 马威 夏梦[1] 郑禄[1] 田望 LEI Jianyun;MA Wei;XIA Meng;ZHENG Lu;TIAN Wang(College of Computer Science&Hubei Provincial Engineering Research Center for Intelligent Management of Manufacturing Enterprises,South-Central Minzu University,Wuhan 430074,China)

机构地区：[1]中南民族大学计算机科学学院&湖北省制造企业智能管理工程技术研究中心,武汉430074

出　　处：《中南民族大学学报（自然科学版）》2023年第6期781-787,共7页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：湖北省科技重大专项(2020AEA011);武汉市科技计划应用基础前沿项目(2020020601012267)。

摘　　要：针对人脸情绪识别类内差异大,类间差异小的特点,结合学生人脸图像的线上课堂情绪识别的场景,提出多尺度空洞卷积模块提取不同空间尺度特征的稠密深度神经网络模型,实现自然场景下学生人脸图像识别.该模型主要由多尺度空洞卷积和DenseNet神经网络两个子网络组成,其中多尺度空洞卷积由不同空洞率的四分支网络提取不同尺度特征,空洞卷积减小特征图尺寸,减少DenseNet内存资源占用;最后在DenseNet网络中结合Adam优化器和中心损失函数.使用稠密网络的旁路连接,加强情绪特征传递和复用.研究结果表明:基于稠密深度神经网络的情绪识别网络模型能够有效提高情绪分类的准确率,模型对预处理后的FER2013+数据集识别准确率达到93.99%,可为线上教学反馈提供技术支持.The characteristics of large intra-class differences and small inter-class differences in facial emotion recognition,combined with the scene of online classroom emotion recognition of student face images,a dense deep neural network model with multi-scale atrous convolution modules to extract features of different spatial scales is proposed,that realize student face image recognition in natural scenes.The model is mainly composed of two sub-networks:Multi-scale atrous convolution and DenseNet neural network.The multi-scale atrous convolution extracts features of different scales by four-branch networks with different atrous rates.Atrous convolution reduces the size of the feature map and reduce the memory resource occupation of DenseNet.Finally,the Adam optimizer and the central loss function are combined in the DenseNet network.The bypass connection of the dense network is used to strengthen the transfer and reuse of emotional features.The research results show that:The emotion recognition network model of the network can effectively improve the accuracy of emotion classification based on dense deep neural network,and the recognition accuracy rate of the model for the preprocessed FER2013+data set reaches 93.99%,which provides technical support for online teaching feedback.

关 键 词：人脸情绪识别 稠密神经网络 空洞卷积 中心损失函数 深度学习优化器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]