基于域适应卷积神经网络的人脸表情识别  

Facial Expression Recognition Based on Convolutional Neural Network with Domain Adaption

作　　者：亢洁[1] 李佳伟 杨思力 KANG Jie;LI Jiawei;YANG Sili(School of Electrical and Information Engineering,Shaanxi University of Science and Technology,Xi'an 710021,China;Air Control Room,Unit 294188,Xi'an 710077,China)

机构地区：[1]陕西科技大学电气与信息工程学院,西安710021 [2]部队航空管制室,西安710077,294188

出　　处：《计算机工程》2019年第12期201-206,共6页Computer Engineering

基　　金：国家自然科学基金(61603233);陕西省自然科学基础研究计划(2017JQ6076)

摘　　要：在利用卷积神经网络进行人脸表情识别时,可借助其他数据集进行辅助训练以应对缺少标记数据的情况,但源域数据库和目标域数据库之间的数据分布差异会影响分类正确率。为此,以AlexNet网络为原型构建基于域适应的卷积神经网络结构。通过引入包含注意力机制的SE模块进行特征重标定,同时利用域适应方法减小领域差异性。在人脸识别公开数据集上的实验结果表明,与AlexNet和GoingDeep等网络相比,该网络能够以较少的参数量获得较高的识别正确率。In facial expression recognition using Convolutional Neural Network(CNN),other data sets can assist in training to deal with the lack of tag data.However,the classification accuracy will be affected by the data distribution differences between the source domain database and the target domain database.To address the problem,this paper constructs a domain adaptation-based convolutional neural network structure modeled after AlexNet.The network introduces a SE module including the attention mechanism for feature re-location,and uses the domain adaption method to reduce the differences between domains.Experimental results on the public data sets of facial expression recognition show that the proposed network can achieve a higher recognition accuracy rate than AlexNet and GoingDeep with fewer parameters.

关 键 词：卷积神经网络 人脸表情识别 数据分布 域适应 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]