基于VGG的人脸表情识别与分类  

Facial expression recognition and classification based on VGG

作　　者：周义飏 ZHOU Yiyang(School of Artificial Intelligence,Beijing Normal University,Beijing 100088,China)

机构地区：[1]北京师范大学人工智能学院,北京100088

出　　处：《智能计算机与应用》2021年第9期35-41,共7页Intelligent Computer and Applications

摘　　要：为了使人脸表情识别更加快速、准确,以满足复杂社会情境中的需求,本文研究了基于深度卷积神经网络的人脸表情识别方法,实现了人脸不同离散表情识别分类。针对现有数据集数据量不足、深度网络计算易出现过拟合现象等问题,本文基于人脸图片关键点进行了剪裁,获得64个子区域,将数据扩充为64倍,以达到数据增强的目的;使用基于VGG-19网络模型的卷积神经网络,对动作单元进行分类与强度计算,使用Sigmoid函数,使网络具备多标签多分类能力,并在VGG-19网络的第四组卷积层之后加入一个加权处理层,提高准确率。结果显示,增强后叠加的人脸表情识别与分类基本能够完成,而引入加权处理层后的准确率则得到了显著提高。In order to make facial expression recognition faster and more accurate to meet the needs in complex social situations,this paper studies the facial expression recognition method based on deep convolutional neural networks to realize the recognition and classification of different discrete facial expressions. First of all,in view of the insufficient amount of data in the existing dataset and the prone to over-fitting in deep network calculations,this paper cuts of the face image to obtain64 sub-regions based on the key points and expand the data to64 times to achieve the goal of data enhancement. Secondly,the convolutional neural network based on the VGG-19 network model is used to classify and calculate the strength of the action unit,and the Sigmoid function is used to make the network have multi-label and multi-classification capabilities. Finally,a weighted processing layer is added after the fourth group of convolutional layers of the VGG-19 network to improve accuracy. The results show that the facial expression recognition and classification after data enhancement can basically be completed,and the accuracy of the weighted processing layer has been significantly improved.

关 键 词：人脸表情 面部动作编码系统 动作单元 卷积神经网络 数据增强 加权处理 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]