基于改进VoVNet的人脸表情识别  

FACIAL EXPRESSION RECOGNITION BASED ON IMPROVED VOVNET

作　　者：李娇[1,2] 王彬彬 周心荆 冉峰[1,2] Li Jiao;Wang Binbin;Zhou Xinjing;Ran Feng(Microelectronics R&D Center,Shanghai University,Shanghai 200000,China;Key Laboratory of Advanced Display and System Application,Shanghai University,Shanghai 200000,China)

机构地区：[1]上海大学微电子研究与开发中心,上海200000 [2]上海大学新型显示技术及应用集成教育部重点实验室,上海200000

出　　处：《计算机应用与软件》2023年第8期155-160,200,共7页Computer Applications and Software

基　　金：国家自然科学基金项目(61674100);新型显示技术及应用集成教育部重点实验室课题项目。

摘　　要：针对实时目标检测网络VoVNet应用于人脸表情识别任务输入尺寸大、不够轻量化、准确率不高等问题,提出改进的VoVNet——EM_VoVNet(EMotion_VoVNet)。EM_VoVNet通过采用组合池化的思想,重设VoVNet的浅层特征提取结构,丰富了浅层特征层。另外,还引入了ECA注意力机制,增强重要特征图的表达能力。采用减少模块个数、通道数的模型优化技术构建网络,使得网络参数仅为1.3×105,具有较好的轻量级特性。在FER2013和RAF-DB数据集上的实验均证明,提出的轻量化EM_VoVNet具有良好的识别性能,有利于模型进一步在嵌入式边缘设备上的部署。Aiming at the problems of large input size,insufficient lightweight and low accuracy of real-time target detection network VoVNet in facial expression recognition tasks,this paper proposes an improved VoVNet called EM_VoVNet(EMotion VoVNet).EM_VoVNet used the idea of combination pooling,and reset the shallow feature extraction structure of VoVNet to enrich the shallow feature layer.In addition,ECA attention mechanism was introduced to enhance the expression ability of important feature maps.The model optimization technology of reducing the number of modules and channels was used to build the network.The network parameters were only 1.3×10^(5),which had good lightweight characteristics.Experiments on both FER2013 and RAF-DB datasets show that the proposed lightweight EM_VoVNet has good recognition performance,which is conducive to the further deployment of the model on embedded edge devices.

关 键 词：表情识别 VoVNet 轻量化 特征提取 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]