Extended fisher criterion based on auto-correlation matrix information

Abstract

Fisher's linear discriminant analysis (FLDA) has been attracting many researchers and practitioners for several decades thanks to its ease of use and low computational cost. However, FLDA implicitly assumes that all the classes share the same covariance: which implies that FLDA might fail when this assumption is not necessarily satisfied. To overcome this problem, we propose a simple extension of FLDA that exploits a detailed covariance structure of every class by utilizing revealed by the class-wise auto-correlation matrices. The proposed method achieves remarkable improvements classification accuracy against FLDA while preserving two major strengths of FLDA: the ease of use and low computational costs. Experimental results with MNIST and other several data sets in UCI machine learning repository demonstrate the effectiveness of our method. © 2012 Springer-Verlag Berlin Heidelberg.