Abstract: Enhanced Independent Component Analysis（EICA） is an unsupervised feature extraction method which is presented based on the overall characteristics，so EICA doesn’t fit to solve such a nonlinear problem as face recognition.Unsupervised Discriminant Projection（UDP） technique is developed for dimensionality reduction of high-dimensional data，and it considers both the local characteristics and non-local characteristics，thus UDP is effective for classification of samples.In this paper，an enhanced unsupervised method is introduced，which has advantages of both EICA and UDP as：（1）it can reflect high-order statistics of samples（2）it is able to discover essential data structure，and obtain a set of effective discriminant projection axis for classification.The experiments on the Yale and FERET databases validate the effectiveness of the proposed method.

Key words: local characteristic, global characteristic, Independent Component Analysis（ICA）, Unsupervised Discriminant Projection (UDP), feature extraction, face recognition

摘要： 增强的独立分量分析（EICA）是一种基于样本整体特征的无监督特征抽取方法，并没有考虑样本的局部特征，因此EICA不利于处理人脸识别这类非线性问题的。无监督鉴别投影技术（UDP）用于高维数据压缩，其基本思想是寻找一组有效的投影方向，使得样本投影后，局部散度最小同时非局部散度最大。UDP同时考虑到样本的局部特征和非局部特征，能够反映样本内在的数据关系，因此UDP能够对样本有效地分类。提出了一种增强的无监督人脸鉴别技术，该方法结合了EICA和UDP的优点，能够：（1）反映样本高阶统计特征；（2）发掘样本内在的几何结构，从而有利于分类。在Yale人脸库和FERET人脸库上的实验验证了该算法的有效性。

关键词: 局部特征, 非局部特征, 独立分量分析, 无监督投影鉴别, 特征抽取, 人脸识别