A wavelet-based method of nearest neighbor pattern classification using scale sequential matching

In this method of pattern classification a wavelet transform is used to extract features from the input signal which are then compared in a scale sequential manner (from coarse to fine) to a trained nearest neighbor codebook. At each scale, possible classification categories are eliminated until only one class is left. We apply this pattern classifier to the problem of fingerprinting post-detection radar pulses and analyze its performance in noise using Monte Carlo simulations. To make our classifier shift invariant, we process the input with an undecimated wavelet transform until the pulse edge is sensed and then start decimating the wavelet coefficients as appropriate to each scale