Linear feature extraction for multiclass problems

Abstract

Motivated by the need for a fast and effective feature extraction method for multiclass problems, a feature extraction method is developed to satisfy two requirements: (1) perform on a class-statistics basis (2) use discriminant information about covariance-difference as well as mean-difference. Experiments show that the new feature extraction method has fulfilled the requirements when the number of training samples is large. Experiments with a small number of training samples were also conducted for showing the limitation of feature extraction.