基于边界异类近邻关系构建的新特征提取方法

A New Feature Extraction Method Based on the Marginal Heterogeneous Neighbor Relationship Construction

特征提取广泛应用于模式识别中。它去除原始样本的冗余信息,提取出有助于样本表示或分类的简洁有用的信息。线性鉴别分析(LDA)属于传统的监督特征提取方法,它旨在寻找最小化类内散度(方差)同时最大化类间散度(方差)的低维线性投影子空间。提出一种新的特征提取方法,旨在改进LDA,该方法在LDA的基础上,增加了每个类的中心点与该类边界异类样本的近邻关系,通过类中心对边界异类样本的排斥,扩大了类与类相互的边距,增强了类的可分性。YaleB人脸数据库和CENPARMI手写阿拉伯数字库中的实验结果,证明了新方法确实能够提高分类效果。

Feature extraction is widely used in pattern recognition.It eliminates redundant information from the original data samples so as to extract useful and concise information in favor of the sample representation or classification method.Linear discriminant analysis（LDA）is a traditional supervised feature extraction,which aims to find a low dimensional linear projection subspace that minimizes the scatter（variance）with the class meanwhile maximizes the scatter（variance）between classes.In this paper,a new feature extraction method is proposed,which aims to modify LDA.On the basis of LDA,the proposed method adds the neighborhood among the central point of each class and the marginal heterogeneous samples,to expand the mutual margin of classes,thus enhances the separabilities among classes.The experimental results on YaleB face database and CENPARMI handwritten numeral database show that the proposed method can improve the classification accuracies.