投审稿平台
投稿指南
下载专区
地  址:北京市海淀区中关村科学院
南路6号中国科学院计算所342号 [地图]
《计算机辅助设计与图形学学报》编辑部
邮政编码:100190
电  话:010-62562491
     010-62600342
订阅信息
ISSN   1003-9775
CN    11-2925/TP
邮发代号:82-456
单  价:80.00元
全年订价:960.00元
订阅电话:010-64017032
在线期刊

多通道运动特征融合的微表情识别方法

佘文祥1,2), 刘斌2), 陶建华1,2)*, 张昊1,2), 吕钊1)
1) (安徽大学多模态认知计算安徽省重点实验室 合肥 230601)2) (中国科学院自动化研究所模式识别国家重点实验室 北京 100190)
分类号: TP391.41 DOI: 10.3724/SP.J.1089.2021.18725
出版年,卷(期):页码: 2021 , 33 ( 9 ): 1457-1465 佘文祥
摘要: 针对已有的微表情识别中由于微表情变化幅度不明显, 导致细微特征容易在学习过程中丢失, 从而使模型的性能受到限制的问题, 提出一种基于运动特征的微表情识别方法. 首先分析变化幅度相对明显的区域对微表情识别的影响, 根据生理学研究对微表情变化相对活跃区域进行局部切割, 并使用并行神经网络分别对局部区域和全局区域提取特征; 然后采用一种能够提取特征级运动信息的运动特征提取模块从空间特征图中学习到运动特征, 将运动特征和空间特征进行聚合, 以减少细微特征的丢失; 最后将局部特征和全局的聚合特征组合成新的混合特征用于微表情识别. 实验结果表明, 在MEGC 2019数据集(包含CASME II, SMIC和SAMM)上, 未加权F1分数和未加权平均召回率的结果分别为81.81%和79.01%, 与MEGC 2019最好的方法相比分别提高了2.96%和0.77%, 该方法具有更好的识别性能.
关键词: 微表情识别; 动作特征提取; 神经网络; 局部特征; 特征聚合
Aggregation of Motion Features of Multiple Paths for Micro-Expression Recogni-tion
She Wenxiang1,2), Liu Bin2), Tao Jianhua1,2)*, Zhang Hao1,2), and Lyu Zhao1)
1) (Anhui Province Key Laboratory of Multimodal Cognitive Computation, Anhui University, Hefei 230601)2) (National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing 100190)
abstract: Since the change of micro-expression is not obvious, subtle features are easily lost in the learning process, which limits the performance of the model. A micro-expression recognition method is proposed based on motion features. Firstly, in order to explore the influence of regions with obvious change on micro-expression recognition, the relatively active regions of micro-expression are locally cut according to physiological research, and the features of local region and global region are extracted respectively by using parallel neural network. When extracting features, a motion feature extraction module which can extract feature level motion information is adopted. The module can learn motion features from spatial feature graph, and aggregate motion features and spatial features to reduce the loss of subtle features. Finally, the local aggregate features and global aggregate features are combined to form a new hybrid feature for micro-expression recognition. Experiments on the MEGC 2019 (including CASME II, SMIC, and SAMM) databases show that the results of unweighted F1 score and unweighted average recall are 81.81% and 79.01%, which are improved by 2.96% and 0.77% respectively compared with the best method of MEGC 2019.
keyword: micro-expression recognition; action feature extraction; neural network; local features; feature aggregation
 
Copyright © 2004《计算机辅助设计与图形学学报》版权所有
电话:010-62600342 传真:010-62562491
E_mail:jcad@ict.ac.cn