Color Indexing by Nonparametric Statistics

Abstract

A method for color indexing is proposed that is based upon nonparametric statistical techniques. Nonparametrics compare the ordinal rankings of sample populations, and maintain their significance when the underlying populations are not Normally distributed. The method differs from previous approaches to color indexing, in that it does not involve histogramming. Principal component analysis is performed to extract the three orthogonal axes of maximum dispersion for a given color signature. These axes are then used to select Lipschitz embeddings to generate sets of scalars that combine all color channel information. These scalar sets are compared against a ranked database of such scalars using the Moses test for variance. On the resulting top matches, the Wilcoxon test of central tendency is applied to yield the best overall match.

The method has been tested extensively on a number of image databases, and has been compared against eight standard histogram methods using four color space transformations. The tests have shown its performance to be competitive with, and in certain cases superior to, the best histogram methods. The technique also shows a greater robustness to noise than all histogram methods, with a noise robustness comparable to that of the more expensive Variable Kernel Density method.