摘要
微表情持续时间短、强度低和局部运动的特点,给其识别带来了极大困难。本文提出了一种新的平均光流方向直方图(MHOOF)描述的微表情识别算法,首先检测人脸稠密关键点并根据关键点坐标和人脸动作编码系统(FACS)将人脸区域划分成13个感兴趣区域(ROI),然后提取选定ROI内相邻两帧之间的HOOF特征来检测微表情序列的峰值帧,最后提取从起始帧到峰值帧这一段图片序列的MHOOF特征进行微表情识别。CASME II微表情库上的实验表明,本文提出的MHOOF特征可有效描述微表情的变化,识别率比两种最优的算法MDMO和Di STLBP-RIP分别提升了5.53%和3.12%。
The recognition of micro-expression has been a great challenge for its three characteristics, i. e. , short duration, low intensity and usually local movements. This paper proposed a novel Mean Histogram of Oriented Optical Flow （MHOOF） feature for micro-expression recognition. First, a set of facial feature landmarks were detected and 13 Regions of Interest （ROIs） were partitioned in facial area based on the landmark coordinates and Facial Action Coding System （FACS） , then the apex frame was detected by HOOF feature extracted in some specific ROIs frame-by-frame. Finally, MHOOF features were extracted from the image sequence that from the onset frame to the apex frame for recognition. The experimental results on the ideal spontaneous micro-expression database, namely, CASME II indicate that the proposed method can describe the changes of micro-expression effectively, and improvements of 5.53% and 3. 12% are achieved when compared to the two state-of-the-art algorithms MDMO and DiSTLBP-RIP respectively.
作者
马浩原
安高云
阮秋琦
MA Hao-yuan1,2, AN Gao-yun1,2, RUAN Qiu-qi1,2(1. Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China;2. Beijing Key Laboratory of Advanced Information Science and Network Technology, Beijing 100044, Chin)
出处
《信号处理》
CSCD
北大核心
2018年第3期279-288,共10页
Journal of Signal Processing
基金
国家自然科学基金(61772067
61472030
61471032
61370127)
中央高校基本科研业务费专项资金(2017JBZ108)
关键词
微表情检测
微表情识别
光流
micro-expression detection
micro-expression recognition
optical flow
