基于统计不相关的无监督鉴别投影映射(UDP)特征提取方法

A Method of Feature Extraction Based on Uncorrelative UDP

基于UDP(Unsupervised Discriminant Projection)特征提取方法的研究,本文提出改进算法:具有统计不相关性的UDP计算方法,并探讨该方法的有效性。这种方法的目的是寻求一组最佳鉴别矢量,使投影变换后的特征空间的非局部散度最大、局部散度最小,从而减小最佳鉴别矢量间的统计相关性。通过在不同人脸库上的仿真实验验证了所提出改进算法在一定程度上优于已有的UDP算法。

This paper presents a modified Unsupervised Discriminant Projection algorithm called Uneorrelative UDP, and application value of this identification method is explained by example. The purpose of the method is to seek a set of best identify vectors ： after the projection transformation, feature space nonlocal scatter is maximum, local scatter is minimum, the statistics correlation between the best identify vectors is reduced. Through the sinmlations in different face libraries, the results verify the effectiveness of the proposed algorithm is improved and is partly superior to the existing UDP algorithm.