基于注意力机制的人脸表情识别迁移学习方法  

Transfer learning method for facial expression recognition based on attention mechanism

作　　者：亢洁[1] 李思禹 KANG Jie;LI Si-yu(School of Electrical and Control Engineering,Shaanxi University of Science and Technology,Xi’an 710021,China)

机构地区：[1]陕西科技大学电气与控制工程学院,陕西西安710021

出　　处：《计算机工程与设计》2021年第3期797-804,共8页Computer Engineering and Design

基　　金：国家留学基金项目(201708615011);陕西省社会发展科技攻关基金项目(2016SF-410);西安市科技计划基金项目(2019216514GXRC001CG002-GXYD1.7)。

摘　　要：针对现有的在人脸表情识别中应用的卷积神经网络结构不够轻量,难以精确提取人脸表情特征,且需要大量表情标记数据等问题,提出一种基于注意力机制的人脸表情识别迁移学习方法。设计一个轻量的网络结构,在其基础上进行特征分组并建立空间增强注意力机制,突出表情特征重点区域,利用迁移学习在目标函数中构造一个基于log-Euclidean距离的损失项来减小迁移学习中源域与目标域之间的相关性差异。在数据集JAFFE和CK+上的实验结果表明,该方法相比其它人脸表情识别方法具有更优的识别能力。To solve the problems that the existing convolutional neural network structure used in facial expression recognition is not lightweight enough to extract facial expression features accurately,and that a large amount of expression labeled data is required,a transfer learning method for facial expression recognition based on attention mechanism was proposed.A lightweight network structure was designed,and feature groups were grouped on the basis of it,afterwards a spatial enhanced attention mechanism was established to highlight the key areas of facial expression features.At the same time,transfer learning was used to construct a loss term based on log-Euclidean distance in the objective function to reduce the correlation difference between the source domain and the target domain.Experimental results on the data sets JAFFE and CK+show that the proposed method has better recognition ability than other facial expression recognition methods.

关 键 词：人脸表情识别 卷积神经网络 注意力机制 特征分组 迁移学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]