基于优化ResNet的人脸表情识别  

A Facial Expression Recognition Algorithm Based on Optimized ResNet

作　　者：徐子凡 程科 袁雪梅 姜元昊 XU Zifan;CHEN Ke;YUAN Xuemei;JIANG Yuanhao(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212100)

机构地区：[1]江苏科技大学计算机学院,镇江212100

出　　处：《计算机与数字工程》2024年第12期3491-3495,3535,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:61976241);镇江市国际合作计划项目(编号:GJ2021008);江苏省研究生创新训练项目(编号:KYCX21_3487)资助。

摘　　要：针对传统人脸表情识别算法准确度不高,网络参数较多的问题,提出一种基于优化残差网络的人脸表情识别算法。算法首先用两个标准卷积层提取人脸表情的浅层特征。然后,利用深度可分离卷积混合通道注意力机制改进残差网络来提取人脸表情的深层特征。最后利用softmax函数对提取的特征分类。在人脸表情识别的公开数据集FER2013和CK+上实验,分别得到分类精度为70.57%和99.28%。实验结果表明该算法表现良好,网络具有较强泛化能力,在复杂情况下对人脸表情能够起到较好的识别效果。Aiming at the problems of low accuracy of traditional facial expression recognition algorithms and many network pa-Iameters,a facial expression recognition algorithm based on optimized residual network is proposed.Firstly,two standard convolu-tion layers are used to extract the shallow features of facial expressions.Then,the depth separable convolution hybrid channel atten-tion mechanism is used to improve the residual network to extract the deep features of facial expressions.Finally,softmax function is used to classify the extracted features.Experiments on the public dataset FER2013 and CK+for facial expression recognition show that the classification accuracy is 70.57%and 99.28%respectively.Experimental results show that the algorithm performs well,the network has strong generalization ability,and can play a good role in facial expression recognition in complex situations.

关 键 词：人脸表情识别 轻量型网络 深度可分离卷积 注意力机制 残差网络 

分 类 号：TN911.73[电子电信—通信与信息系统]