Whiteness as a color-classification problem

Abstract

Colors called "white' are found in a small but yet not sharply defined region of the color space. The main characteristics of a white sample are high ligthness, tristimulus value Y > 70 and low saturation, excitation purity p <jo.1 and it is usually assumed that the reflectance spectra are continuous. In our study we measured 24 white samples, Ciba Geigy Plastic White Scale and white ceramic plates, with an acousto-optic color spectrometer and tested several whiteness formulas, based both to conventional color coordinate definitions and to statistical pattern recognition and classification methods, to find out the best whiteness formula for quality inspection and assortment. Each sample was measured 20 times in the same geometry (0/45) using normal tungsten halogen lamp as a light source, and the reflectance spectrum was measured from 400 nm to 700 nm with 5 nm intervals in order to achieve high accuracy for color determination.The aim was to classify measured white data to preknown classes using different decision rules, whiteness formulas. Tristimulus based formula showed classification power of 40-53 % and color coordinate based formula showed corresponding values. Using spectral information of the samples the classification power was improved clearly, with a sirriple difference algorithm it was 76 % . The low values of classification powers in general are due the "difficult" data of ceramic plates and direct comparision of the measured spectra showed that there were almost identical plates, (preceived identical by human observers). The decision rules, that used only chromatic information, showed to be improper, because a large set of spectra is almost identical in shape but having minor differences in lightness. In two class problem it was always possible to find a decision rule for exact classification using-chromatic informat.oi only e.g. by unified pseudoinverse algorithm or by subspace approach,' but extensions to a larger data set were easier if a measure of lightness was taken into account to reduce the number of classes back to two or three. Optical realizations of subspace classifiers include acousto-optic colorimeter with adjustable diffraction efficiency and use of 2 D-spatial light modulators to manipulate the spectrum of dispersed object light after a grating or a prism.This work was supportedby the Academy of Finland 1.