基于特征提取和胶囊网络的人脸表情识别  

Facial expression recognition based on feature extraction and capsule network

作　　者：黄小刚 黄润才[1] 王桂江 马诗语 HUANG Xiaogang;HUANG Runcai;WANG Guijiang;MA Shiyu(School of Electronic and Electrical Engineering,Shanghai University of Engineering Science,Shanghai 201620,China)

机构地区：[1]上海工程技术大学电子电气工程学院,上海201620

出　　处：《智能计算机与应用》2022年第10期68-74,共7页Intelligent Computer and Applications

摘　　要：人脸表情识别是近年来非常火热的一个研究领域,随着深度学习的发展,越来越多的深度学习方法用于表情识别中。针对胶囊神经网络(CapsNet)更关注的是图像高层空间信息、低层空间特征提取不全面的问题,提出了特征提取与胶囊网络结合的人脸表情识别算法。本文先使用局部二值模式(LBP)算子提取图像纹理特征,与胶囊网络结合形成多通道输入胶囊网络。为了进一步加强低层空间特征提取,在提取纹理特征后加入了深度残差网络(ResNet),与胶囊网络结合形成多通道输入增强胶囊网络。为了验证多通道输入胶囊网络和多通道输入增强胶囊网络的性能,本文在公开表情数据集CK+和RAF-DB分别进行了对照实验,得到了99.69%,82.02%准确率,优于其它的表情识别算法。Facial expression recognition is a very hot research field in recent years.With the development of deep learning,more and more deep learning methods are used in expression recognition.Aiming at the problem that capsule neural network(CapsNet)pays more attention to the high-level spatial information of the image,but the low-level spatial feature extraction is not comprehensive,a facial expression recognition algorithm combining feature extraction and capsule network is proposed.Firstly,local binary pattern(LBP)operator is used in this paper to extract image texture features,and combined with capsule network,multi-channel input capsule network is formed.After that,in order to further strengthen low-level spatial feature extraction,depth residual network(ResNet)is added after extracting texture features.Finally,combined with capsule network,multi-channel input enhanced capsule network is formed.Based on the above,this paper carries out control experiments on the public expression data sets CK+and RAF-DB respectively,which is used to verify the performance of the multi-channel input capsule network and the multi-channel input enhanced capsule network.The simulation achieves 99.69%and 82.02%accuracy,respectively and the results proves that the proposed method in this paper is better than other expression recognition algorithms.

关 键 词：人脸表情识别 特征提取 深度残差网络 胶囊网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]