基于递增注意力的微表情识别方法  

Micro-expression recognition method based on progressive attention

作　　者：战子为 孙兆才 李翔 吴镇东 ZHAN Ziwei;SUN Zhaocai;LI Xiang;WU Zhendong(Qingdao Academy of Traditional Chinese Medicine,Shandong University of Traditional Chinese Medicine,Qingdao 266112,Shandong Province,P.R.China;Center for Medical Artificial Intelligence,Shandong University of Traditional Chinese Medicine,Qingdao 266112,Shandong Province,P.R.China;Qingdao Key Laboratory of Traditional Chinese Medicine Artificial Intelligence Technology,Shandong University of Traditional Chinese Medicine,Qingdao 266112,Shandong Province,P.R.China)

机构地区：[1]山东中医药大学青岛中医药科学院,山东青岛266112 [2]山东中医药大学医学人工智能研究中心,山东青岛266112 [3]山东中医药大学青岛市中医人工智能技术重点实验室,山东青岛266112

出　　处：《深圳大学学报（理工版）》2024年第6期756-764,I0011,共10页Journal of Shenzhen University(Science and Engineering)

基　　金：国家自然科学基金资助项目(62372280,61872225);山东省自然科学基金资助项目(ZR2020KF013,ZR2020QF043,ZR2023QF094);青岛市科技惠民示范专项资助项目(23-2-8-smjk-2-nsh)。

摘　　要：微表情是个体无意识发生的表情变化,能够反映人们潜在的情绪和内心状态.微表情发生时动作强度低且涉及面部范围小,导致在微表情识别过程中存在着提取特征不充分、定位有效特征不准确的问题,影响识别精度.构建一种递增注意力多尺度卷积网络,该网络融合了多尺度卷积模块和递增注意力模块.利用多尺度卷积模块学习不同感受野下的细粒度特征,提取丰富的细节特征,同时设计一种递增注意力模块,通过多个注意力图间的特征共享与增强,准确定位面部运动区域,稳健提取微表情图像中的运动特征.所提网络在数据集SMIC、CASMEII及SAMM上进行实验,准确率分别达到0.826、0.880和0.787,F1值分别达到0.817、0.864和0.761.研究结果可为谎言检测、心理健康早期筛查等提供参考.Micro-expression is unconscious expression changes that reflect people's underlying emotions and inner states.When micro-expressions occur,their low intensity and the small facial range result in insufficient feature extraction and inaccurate localization of effective features during the recognition process,which affects recognition accuracy.To address this issue,a progressive attention multi-scale convolutional network was constructed.The network integrates a multi-scale convolutional module and a progressive attention module.First,the multi-scale convolutional module is used to learn fine-grained features from different receptive fields,extracting rich details.Then,the progressive attention module is designed to accurately locate facial motion areas and robustly extract motion features from micro-expression images through information sharing and enhancement across multiple attention maps.The proposed network was tested on SMIC,CASMEII and SAMM datasets,achieving accuracy rates of 0.826,0.880 and 0.787,and F1 scores of 0.817,0.864 and 0.761,respectively.The proposed method can serve as an auxiliary tool for lie detection and early screening of mental health conditions.

关 键 词：人工智能 微表情识别 深度学习 注意力机制 卷积神经网络 多尺度卷积 谎言检测 心理健康早期筛查 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] R318[自动化与计算机技术—计算机科学与技术]