大型人脸库上基于局部二元模式直方图映射的快速识别

Local binary pattern histogram projection based fast recognition on large-scale face database

传统的基于局部二元模式(LBP)的人脸识别方法采用卡方统计度量LBP直方图间的差异,由于卡方统计度量的复杂性以及是在高维空间进行判别,此方法在大型人脸库上的识别速度低,为此提出一种LBP直方图映射(LBPHP)方法.将LBP直方图映射到保局投影(LPP)空间获取低维LBPHP特征,当判别新样本时只须比较新样本与训练样本的LBPHP特征,识别过程简单且在低维空间进行,识别速度很快.鉴于LPP强大的鉴别特性,此方法的识别率很高.在2个知名人脸库上对LBPHP方法进行实验验证,结果表明,相比于传统识别方法,LBPHP的识别速度快,尤其在大型人脸库上优势更加明显,适于在此类人脸库上的实际应用如身份认证等.

The conventional local binary pattern （LBP） based facial recognition method selects Chi square statistic as the dissimilarity measure for LBP histogram.In view of Chi square statistic＇s complexity and the high-dimensional recognition process,the conventional method is very slow as recognizing on large-scale face database.A method of LBP histogram projection （LBPHP） was proposed,which projects LBP histogram onto locality preserving projection （LPP） space and obtains low-dimensional LBPHP feature.Recognizing new sample only needs to compare its LBPHP feature with those of training samples.The process is simple and carried on low-dimensional space,thus the proposed method is fast and has good accuracy in the light of powerful discriminative property of LPP.Comparative experiments on two large-scale face databases demonstrated that the LBPHP method is superior to the conventional method on recognition speed.The LBPHP method is prominent especially on large-scale face database and suitable for practical application,e.g.identity authentication.