基于改进MobileNet网络的人脸表情识别  

FACIAL EXPRESSION RECOGNITION BASED ON IMPROVED MOBILENET

作　　者：王韦祥 周欣 何小海[1] 卿粼波[1] 王正勇[1] Wang Weixiang;Zhou Xin;He Xiaohai;Qing Linbo;Wang Zhengyong(College of Electronics and Information Engineering,Sichuan University,Chengdu 610065,Sichuan,China;China Information Technology Security Evaluation Center,Beijing 100085,China)

机构地区：[1]四川大学电子信息学院,四川成都610065 [2]中国信息安全测评中心,北京100085

出　　处：《计算机应用与软件》2020年第4期137-144,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61871278);四川省科技计划项目(2018HH0143);四川省教育厅项目(18ZB0355);成都市产业集群协同创新项目(2016-XT00-00015-GX)。

摘　　要：针对轻量级卷积神经网络MobileNet应用于人脸表情识别实时性较差、最小输入尺寸较大、准确率不高等问题,提出一种改进的MobileNet网络模型——M-MobileNet(Modified MobileNet)。M-MobileNet具有比原网络更好的轻量级特性。该网络模型基于一种改进的深度可分离卷积层,不仅具有MobileNet模型中深度可分离卷积减少卷积计算量的特点,还解决了在深度卷积层后可能会导致信息丢失的问题。在分类器选择上,M-MobileNet使用线性支持向量机(SVM)进行人脸表情分类,参数量较MobileNet网络大大减少。在CK+、KDEF数据集及移动端上的实验证明,改进后的MobileNet网络模型具有更好的识别性能。In order to solve the problems of poor real-time performance,large minimum input size and low accuracy of lightweight convolutional neural network MobileNet in facial expression recognition,we propose an improved MobileNet network model called M-MobileNet(Modified MobileNet).It has better lightweight characteristics than the original network.M-MobileNet is based on an improved depthwise separable convolution layer.It has the characteristics of depthwise separable convolution in MobileNet model to reduce the amount of convolution calculations,and can the problem that information may be lost after depthwise convolution layer.The M-MobileNet network model used the linear support vector machine(SVM)to classify facial expressions,and the parameters of M-MobileNet were greatly reduced,compared with the MobileNet network.The experiments on CK+and KDEF data sets and mobile terminals show that the improved MobileNet network model has better recognition performance.

关 键 词：MobileNet 表情识别 深度可分离卷积 支持向量机 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]