Kernel Fisher Discriminant Analysis with Locality Preserving for Feature Extraction and Recognition
Di Zhang, Jiazhong He, Yun Zhao · International Journal of Computational Intelligence Systems · 2013
Many previous studies have shown that class classification can be greatly improved by kernel Fisher discriminant analysis (KDA) technique.However, KDA only captures global geometrical structure and disregards local geometrical structure of the data.In this paper, we propose a new feature extraction algorithm, called locality preserving KDA (LPKDA) algorithm.LPKDA first casts KDA as a least squares problem in the kernel space and then explicitly incorporates the local geometrical structure information into the least squares problem via regularization technique.The fact that LPKDA can make full use of two kinds of discriminant information, global and local, makes it a more powerful discriminator.Experimental results on four image databases show that LPKDA outperforms other kernel-based algorithms.