一种基于光流双输入网络的微表情顶点帧检测方法  

A Micro-Expression Apex Frame Spotting Method Based on Optical-Flow-Dual-Input Network

作　　者：郑戍华[1] 陈梦心 王向周[1] 弓雪雅 ZHENG Shuhua;CHEN Mengxin;WANG Xiangzhou;GONG Xueya(School of Automation,Beijing Institute of Technology,Beijing 100081,China)

机构地区：[1]北京理工大学自动化学院,北京100081

出　　处：《北京理工大学学报》2022年第7期749-754,共6页Transactions of Beijing Institute of Technology

基　　金：国家部委预研资助项目(5200-2020036147A-0-0-00)。

摘　　要：微表情顶点帧蕴含着丰富的微表情信息,为了准确地检测出微表情顶点帧,本文提出了一种基于光流特征的神经网络分类,并利用先验知识规则进行取舍的检测方法.该方法针对固定滑窗大小内的图像进行光流信息提取,利用双输入特征提取网络对x,y方向的光流信息进行时空特征提取,并进行分类,经根据微表情先验知识所设计的取舍规则后处理后,改善了检测准确度.实验结果表明,在数据集CASMEⅡ上测试,顶点定位率(apex spotting rate,ASR)指标达到了0.945,F_(1)-score指标达到了0.925.Micro-expression apex frame contains abundant micro-expression information.In order to spot the apex frame accurately,a neural network classification was proposed based on optical flow characteristics.Taking prior knowledge as rules,a detection method was designed to realize micro-expression apex frame spotting.Firstly,optical flow information was extracted from the image in a fixed size sliding window.And then,the spatial and temporal features of optical flow information in x and y directions was extracted and classified based on dual input network.Finally,according to the trade-off rules based on prior knowledge of micro expression,a post-processing was carried out to improve the detection accuracy.The experimental results on data set CASMEⅡtesting show that the apex spotting rate(ASR)and F_(1)-score can reach up to 0.945 and 0.925 respectively.

关 键 词：微表情顶点帧 双输入网络 分类后处理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]