基于面部关键点和图卷积的表情识别方法  

Facial Expression Recognition Method Based on Facial Landmark and Graph Convolution

作　　者：吴宇凡 李春国[1] 杨绿溪[1] WU Yufan;LI Chunguo;YANG Lvxi(School of Information Science and Engineering,Southeast University,Nanjing 211189,China)

机构地区：[1]东南大学信息科学与工程学院,江苏南京211189

出　　处：《无线电通信技术》2022年第5期924-929,共6页Radio Communications Technology

基　　金：国家自然科学基金(62171119);江苏省重点研发计划(BE2021013-3)。

摘　　要：针对现有人脸表情识别方法对于面部细节处的局部特征关注度不足的问题,提出了基于面部关键点和图卷积的人脸表情识别方法CGNet。CGNet将面部图像按面部器官进行分割得到多个分割图像,提取分割图像的多尺度特征并引入空间注意力机制提取细节信息,提升网络对于面部细节的关注度;提取人脸关键点,利用图卷积网络提取出人脸面部的结构信息,提升网络对高维度特征的表示能力。实验结果表明,CGNet是一种高效的表情识别算法,能够获得更有效的面部特征,提高识别准确率。Various methods on facial expression recognition have been proposed, where local features of facial details haven’t got sufficient attention.To solve this problem, we propose a facial expression recognition method called CGNet which is based on facial landmark points and graph convolutional network.Facial images are segmented according to facial organs and then input into the network as multi-scale information.Spatial attention mechanism is applied in CGNet in order to extract detailed information and raise the attention on facial details of the network.By extracting facial landmark points and facial structure information with graph convolutional network, CGNet improves its ability to represent high-dimensional features.Experiments demonstrate that CGNet is an efficient facial expression recognition method which is able to achieve more useful facial features and improve recognition accuracy.

关 键 词：人脸表情识别 面部关键点 图卷积 空间注意力 多尺度特征 

分 类 号：TN919.23[电子电信—通信与信息系统]