Investigation of Feature Selection Criteria for Pattern Recognition Models Including the Fourier Transmission.
AIR FORCE INST OF TECH WRIGHT-PATTERSON AFB OH SCHOOL OF ENGINEERING
Pagination or Media Count:
Feature selection is of fundamental importance in pattern recognition. The investigation evaluates and compares 10 feature selection criteria. The two-dimensional, discrete Fourier transform is specified so that the low-pass spatial filter criterion can be included in the comparison. Feature space extraction and feature space evaluation processes are modeled and implemented. Two sets of data consisting of handprinted characters are used in a series of experiments that extract feature spaces corresponding to the various criteria and evaluate the feature spaces by a class separability measure and an error estimate. The results are tabulated for comparison and conclusions are drawn on the empirical and theoretic bases established. Modified author abstract