基于改进的Fisher准则的多示例学习视频人脸识别算法  

Video Face Recognition Based on Modified Fisher Criteria and Multi-instance Learning

作　　者：王玉[1,2,3] 申铉京[1,3] 陈海鹏[1,3] WANG Yu;SHEN Xuan-Jing;CHEN Hai-Peng(College of Computer Science and Technology,Jilin University,Changchun 130012;Applied Technology College,Jilin University,Changchun 130012;Key Laboratory of Symbolic Computation and Knowledge Engineering of Ministry of Education,Jilin University,Changchun 130012)

机构地区：[1]吉林大学计算机科学与技术学院,长春130012 [2]吉林大学应用技术学院,长春130012 [3]吉林大学符号计算与知识工程教育部重点实验室,长春130012

出　　处：《自动化学报》2018年第12期2179-2187,共9页Acta Automatica Sinica

基　　金：国家青年科学基金(61305046;61602203);吉林省优秀青年人才基金(20180520020JH)资助~~

摘　　要：视频环境下目标的姿态变化使得人脸关键帧难以准确定位,导致基于关键帧标识的视频人脸识别方法的识别率偏低.为解决上述问题,本文提出一种基于Fisher加权准则的多示例学习视频人脸识别算法.该算法将视频人脸识别视为一个多示例问题,将视频中归一化后的人脸帧图像作为视频包中的示例,采用分块TPLBP级联直方图作为示例纹理特征,示例特征的权值通过改进的Fisher准则获得.在训练集合的示例特征空间中,采用多示例学习算法生成分类器,进而实现对测试视频的分类及预测.通过在Honda/UCSD视频库和Youtube Face数据库中的相关实验,该算法达到了较高的识别精度,从而验证了算法的有效性.同时,该方法对均匀光照变化、姿态变化等具有良好的鲁棒性.Due to the pose variation of target in video,it is difficult to accurately locate the face key frame and have a high recognition rate of the video face recognition based on key frame identification.To solve these problems,a video face recognition algorithm based on multi-instance learning is proposed in this paper.The algorithm takes each face video as a bag,and each normalized face frame as an instance in the bag.The feature of each instance is represented by cascading histograms of block TPLBP codes,and the weight of the instance feature is obtained by the improved Fisher criteria. The classifier is obtained in the feature space of training set by using a multiple instance learning algorithm,and then classification and prediction of test bag are realized.Experiments on the Honda/UCSD and YouTube Face databases show that the algorithm can achieve a higher recognition accuracy,and at the same time,the method is robust to illumination variation and expression variation.

关 键 词：视频人脸识别 局部二值模式 多示例学习 FISHER准则 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]