人脸微表情分析方法综述  

A Survey of Facial Micro-Expression Analysis Methods

作　　者：于明 钟元想 王岩[2] YU Ming;ZHONG Yuanxiang;WANG Yan(School of Artificial Intelligence,Hebei University of Technology,Tianjin 300401,China;School of Information Engineering,Tianjin University of Commerce,Tianjin 300134,China)

机构地区：[1]河北工业大学人工智能与数据科学学院,天津300401 [2]天津商业大学信息工程学院,天津300134

出　　处：《计算机工程》2023年第2期1-14,共14页Computer Engineering

基　　金：河北省自然科学基金(F2019202381,F2019202464,F2020202025,F2021202030)。

摘　　要：微表情分析在医学、公共安全、商业谈判等领域得到广泛应用并备受关注。微表情运动幅度小、变化快,导致人工分析难度较大,开发一个可靠的自动化微表情分析系统非常有必要。随着计算机视觉技术的发展,研究人员能够结合相关算法捕捉微表情运动变化特征以用于微表情分析。阐述微表情分析的发展历程和现状,从多个角度对微表情分析的两大分支,即微表情检测方法和微表情识别方法进行总结。整理现有微表情数据集以及微表情分析流程中常用的面部图像预处理方法。根据特征提取方式的不同,从基于时间特征、基于特征变化和基于深度特征这3个方面对微表情检测方法进行阐述。将微表情识别方法归纳为基于纹理特征和基于光流特征的传统机器学习方法以及深度学习方法,其中,基于深度学习的微表情识别包括基于运动单元、基于关键帧和基于迁移学习的方法。通过不同实验指标对以上方法进行分析和比较,在此基础上,探讨当前微表情分析中存在的问题和挑战并展望该领域未来的发展方向。Micro-expression analysis has been widely used in the medical industry,public safety,business disputes and other fields and has attracted much attention.Manual analysis is challenging because the intensity of micro-expressions is relatively modest and they fluctuate quickly.Therefore,automated methods must be developed to perform microexpression analysis.Driven by advancements in learning models,many works on computer vision have considered the use of subtle facial movements to investigate micro-expressions.This study thoroughly describes the evolution of microexpression analysis.The two key issues involved,including detection and identification methods,are summarized rigorously from various perspectives in light of the current state of the art.First,existing micro-expression datasets and facial image preprocessing methods that are commonly used in micro-expression analysis are summarized and introduced.Then,according to different feature extraction methods,micro-expression detection methods are divided into those that consider temporal features,feature change,and depth features.Micro-expression recognition techniques are largely based on texture features,optical flow features of traditional machine learning methods,and deep learning methods.Deep learning methods have been developed for micro-expression recognition based on motion units,key frames,and transfer learning methods.Thus,several experimental indicators are analyzed and compared.In conclusion,the current problems and challenges in micro-expression analysis are discussed and some possible directions for future development are suggested.

关 键 词：微表情分析 计算机视觉 微表情检测 微表情识别 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]