相关NPE算法的人脸识别研究  

Facial recognition research on correlation NPE algorithm

作　　者：刘嘉敏[1] 李连泽 罗甫林[1] 刘亦哲[1] 刘玉梅[1] 

机构地区：[1]重庆大学光电技术及系统教育部重点实验室,重庆400044

出　　处：《计算机应用研究》2015年第6期1881-1883,1891,共4页Application Research of Computers

基　　金：中央高校基本科研业务费资助项目(1061120131207;CDJXS12120001)

摘　　要：传统的近邻保持嵌入(NPE)算法采用欧氏距离作为近邻点选取的度量,但欧氏距离只表示两点间的直线距离,在高维空间中不一定能反映数据间的真实空间分布,易导致近邻选取不准确。针对此问题,提出了相关近邻NPE(CNPE)算法。该方法利用相关系数度量数据间的近邻关系,实现更准确的局部重构,提取更有效的鉴别特征。在CMU PIE人脸数据集上的实验结果表明,提出的CNPE算法比NPE、LLE、LPP拥有更高的识别率。CNPE算法增加了近邻为同类的概率,能更有效地实现人脸识别。Traditional NPE algorithm considers Euclidean distance as the standard to select neighbor points, but the Euclidean distance only represents the linear distance between two points, it couldn＇ t reflect data＇ s reality distribution in the high-dimen- sional space which caused instability in neighbor points selecting. Aiming at this problem, this paper put forward CNPE algo- rithm. It utilized correlation coefficient to weigh neighbor relationship among data, which realized more accurate local recon- struction and extracted more abundant discriminate features. The experiments on CMU PIE face database reveals that the CNPE algorithm, this paper proposed owns higher recognition rate than NPE, LLE and LPP algorithm. What＇ s more, the algorithm in- creases the possibility that the neighbor points to be the homogeneous points, which could achieve face recognition efficiently.

关 键 词：人脸识别 流形学习 近邻保持嵌入 相关系数 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]