The role of spectral decomposition in the pattern recognition of narrowband signals

The use of spectrum estimators as preprocessors to classification decisions is discussed in this paper. The classification performance using features chosen after spectrum estimation is measured by estimating the Bayes error, obtained by using the kth Nearest Neighbor (kNN) algorithm. Two spectrum estimators, are used to preprocess three different narrowband signals immersed in additive noise and classification comparisons made. The paper concludes that, for the techniques and features used here-in, classification performance is not significantly affected by the spectral estimation methods employed to form the features.