基于旋转不变局部相位量化特征的人脸确认算法研究  

Search of face verification algorithm based on RILPQ descriptor

作　　者：高志升[1] 袁红照[1] 杨军[1] 

机构地区：[1]四川大学图形图像研究所,成都610064

出　　处：《计算机应用研究》2012年第1期352-354,362,共4页Application Research of Computers

基　　金：国家"863"计划资助项目(2006AA12A104)

摘　　要：针对非受控环境下人脸图像的采集易受光照、姿态、表情、遮挡的影响且成像质量低等为人脸确认带来很大困难这一问题,提出了采用旋转不变局部相位量化(RILPQ)特征算子结合学习度量距离的方法进行人脸确认。首先利用RILPQ特征算子对待确认的两幅图像分别提取RILPQ编码图像;然后分块获得空间区域RILPQ直方图序列并进行PCA降维,并将降维后的RILPQ直方图序列作为人脸图像的特征描述子,计算两幅人脸图像描述特征的统计Fisher加权距离;最后采用SVM进行人脸确认。在LFW人脸库上的实验表明该方法在同类算法中具有最好的性能。The problem that face image captured from uncontrolled environment is seriously affected by illumination,pose,expression,occlusion etc.and has low image quality has brought about much difficulty in face verification.To solve this problem,this paper proposed a novel method for face verification which combining rotation invariant local phase quantization（RILPQ） with learning metric distance.First,extracted the RILPQ code map from two given face images respectively by using RILPQ descriptor.Then,divided the code map into N×N regions,after concatenating all RILPQ histograms from every region and PCA dimension reduction,and represented each given face image by the compact RILPQ feature.Next,obtained similarity between the two face images by using statistical Fisher weight distance.And finally,adopted SVM for face verification.Experiments on LFW database show that the method achieves best performance among all the same kind of methods.

关 键 词：人脸确认 RILPQ 主成分分析 统计Fisher加权距离 支持向量机 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]