基于面向分类准则的维数约简及其在人脸识别中的应用  

Classification Oriented Criterion Based Dimensionality Reduction and its Application in Face Recognition

作　　者：殷飞[1] 焦李成[1] 

机构地区：[1]西安电子科技大学智能感知与图像理解教育部重点实验室,西安710071

出　　处：《计算机科学》2014年第5期283-287,共5页Computer Science

基　　金：国家重点基础研究发展计划(2013CB329402);国家自然科学基金(61072106;61173090;61072108;61303034);教育部长江学者和创新团队发展计划(IRT1170);基本科研业务费(K5051303011)资助

摘　　要：针对高维数据导致的维数灾难问题,提出了一种基于面向分类准则的维数约简方法。所提准则使每个训练样本在特征空间中与同类样本尽可能接近,而与异类样本尽可能疏远。首先对每个训练样本定义同类样本加权平均距离和异类样本加权平均距离。然后基于上述两个概念分别定义总体同类距离和总体异类距离。以最小化总体同类距离和最大化总体异类距离为目的提出了面向分类的准则(Classification Oriented Criterion,COC)。最后,基于面向分类的准则推导出了一种新的维数约简方法。在公共人脸数据库ORL和Yale上的实验表明所提方法性能优于有代表性的维数约简方法。To tackle the problem of the curse of dimensionality caused by high dimensional data,a classification oriented criterion based dimensionality reduction method was presented.The proposed criterion aims at making each training sample and samples from the same class as close as possible in the feature space and making each training sample and samples from the different classes as distant as possible in the feature space.First,for each training sample,weighted average distance of samples from the same class and weighted average distance of samples from different classes were defined.Then,based on these two concepts,total distance of samples from the same class and total distance of samples from different classes were defined.After that,Classification Oriented Criterion （COC） was proposed,which aims at minimizing the total distance of samples from the same class and maximizing the total distance of samples from different classes.Finally,a novel dimensionality reduction method based on COC was presented.The experiments on publicly available face databases ORL and Yale demonstrate that the proposed method outperforms representative dimensionality reduction methods.

关 键 词：维数约简 总体同类距离 总体异类距离 面向分类的准则 人脸识别 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]