基于光流和自编码器的微表情检测方法  

MICRO-EXPRESSION DETECTION METHOD BASED ONOPTICAL FLOW AND AUTOENCODER

作　　者：黄树成[1] 罗德广 Huang Shucheng;Luo Deguang(School of Computer,Jiangsu University of Science and Technology,Zhenjiang 212003,Jiangsu,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212003

出　　处：《计算机应用与软件》2023年第9期171-176,共6页Computer Applications and Software

基　　金：国家自然科学基金项目(61772244)。

摘　　要：微表情检测广泛应用在谎言识别、心理健康和情感分析等场合,构建微表情检测模型需要充足的训练数据,但是标注微表情需要过高的成本,导致自发性微表情样本库数量过少,给微表情检测带来了极大的挑战。针对这个挑战提出一种新的微表情检测方法FLOW-AENET:提取人脸的光流特征,将光流特征作为自编码器的输入,利用深度学习模型对特征进行处理,再将学习到的特征加入SVM分类器中做二分类,在含有微表情的一类中,根据ROIS区域的变化程度判断出微表情产生的起始帧、顶峰帧和结束帧。在CASEME、CASME II等数据集上进行实验研究,结果表明,FLOW-AENET方法相比于其他方法具有明显的优势。Micro-expression detection is widely-used in lies recognition,mental health and sentiment analysis.Building a micro-expression detection model requires sufficient training data.However,labeling micro-expressions requires excessively high costs,resulting in a small number of spontaneous micro-expression samples,which brings great challenges to micro-expression detection.Aiming at this challenge,this paper proposes a new micro-expression detection method FLOW-AENET.The optical flow features of human face were extracted,and the optical flow features were used as the input of the autoencoder.The deep learning model was used to process the features,and the features were added to the SVM classifier for secondary classification.In the category containing micro-expression,the start frame,peak frame and end frame of the micro-expression were determined by the degree of change of the ROIS region.The FLOW-AENET method was experimentally studied on CASEME,CASME II and other datasets.The results show that it has obvious advantages compared with other methods.

关 键 词：微表情检测 光流特征 自编码器 SVM分类器 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]