The effect of histogram discontinuities on spectral information and non-supervised classifiers

Abstract

Discontinuities in histograms may appear as side effects in the acquisition and pre-processing of images. Although these can be indistinguishable from real cluster peaks in univariate histograms, they are clearly detected in bivariate histograms. The effect of such discontinuities over the isodata algorithm and over a non-parametric clustering algorithm has been tested. Though it is not clear that they directly affect the isodata classifier, classes created do not match at all the number, shape and location of spectral clusters observed in bivariate histograms. Conversely, the non-parametric clustering algorithm can be clearly affected by histogram discontinuities; but, if these are previously removed by a low pass filter, it may be able to describe correctly the structure of the spectral space.