基于核邻域保持判别嵌入的人脸识别

Face recognition based kernel neighborhood preserving discriminant embedding

为了提取高维人脸图像中的非线性特征,提出一种新的非线性降维方法:核邻域保持判别嵌入算法(KNP-DE).为了表示特征空间中类间邻域结构和不同类样本间的相似度,分别构建类内邻接矩阵和类间相似度矩阵.通过使用核技巧,KNPDE将邻域保持嵌入(NPE)和Fisher判别准则相结合,在保持特征空间中类内邻域结构的同时充分利用类间判别信息,从而具有更强的分类能力.在Yale和UMIST人脸库上的试验结果进一步表明了该算法的有效性.

A novel nonlinear dimensionality reduction method named kernel neighborhood preserving discriminant embedding （KNPDE） was proposed in order to extract nonlinear feature in high dimensional face image. The within-class affinity matrix and the between-class similarity matrix were constructed respectively in order to represent the within-class neighborhood geometry and the similarity between the samples from different classes in feature space. KNPDE integrated neighborhood preserving embedding （NPE） with Fisher discriminant criterion by using kernel trick. KNPDE possessed much more power in classification, which preserve the within-class neighborhood geometry in feature space and sufficiently use the between-class discriminant information. Experimental results on the Yale and the UMIST face databases demonstrated the effectiveness of the algorithm.