: Pannekamp, J.; Schmutz, M.

Abstract
In this paper a methodology for selecting features for the multiclass-problem is discussed. The approach is based on an iterative solution scheme. The two major steps "selection of feature subsets" and "evaluation of their effectiveness" are realized as a sequential forward search strategy combined with the evaluation of the Kullback-Leibler distance in the approximation for normal distributions. Thus, the method becomes computationally very efficient and may be applied to feature pools of up to several hundred primary features. Its effectiveness is demonstrated for a task from the field of texture segmentation.