摘要：

监督的保持邻域嵌入算法采用欧氏度量选取k近邻。欧氏度量在数据维数较低时能获得较好的结果，但直接简单地将其从低维空间的应用推广到高维空间中不能取得较好的结果。针对该缺点，提出度量优化的保持邻域嵌入算法。该算法分为无类标号信息(MONPE)和有类标号信息(CLMONPE)2种情况，利用线性判别分析算法降维后的数据选取k近邻。在Yale人脸数据库上的实验结果表明，CLMONPE算法效果较优。

关键词: 流形学习, 人脸识别, 监督的保持邻域嵌入, 度量优化的保持邻域嵌入

Abstract:

Euclidean metric is adopted to look for k-nearest neighbors in the supervised Neighborhood Preserving Embedding(NPE). However, the results are not very good when Euclidean metric is directly generalized to handle high-dimensional data as dealing with low-dimensional data. To overcome this problem a metric-optimized neighborhood preserving embedding algorithm is proposed in this paper. Two conditions are considered: non-labeled case(MONPE) and labeled case(CLMONPE). The main idea is to choose k-nearest neighbors by analyzing the data whose dimension is reduced with linear discriminant analysis algorithm. Test result on Yale database shows that CLMONPE has obvious strength in application.

Key words: manifold learning, face recognition, supervised Neighborhood Preserving Embedding(NPE), metric-optimized NPE

中图分类号: 

• TP18