The use of color in computational vision

A method, based on finite-dimensional linear models of reflectance and illumination, which allows the transformation of images into color constant images is used, and it is shown that good results can be obtained using a 3-receptor system and statistical measurements of natural materials and illuminants.

Abstract

Color has always been part of the visual percept, yet research in computer vision seemed to focus on other properties, mainly because they did not require the additional information color provides. In this thesis, different aspects of color perception and computation are discussed. We start by examining and analyzing biologically motivated models of early chromatic visual processing. Two non-linear models of color measurement are investigated, and tested with different stimuli designed to reveal some of their spatio-chromatic properties. This analysis leads to several speculations about the use and functionality of the models and the operators they employ. One application of the operators previously discussed is to the problem of distinguishing shadow boundaries from material changes. We examine and formulate the behavior of shadows under different illumination conditions through the use of a model of reflection. Based on this analysis, we suggest a technique which makes use of a subset of the operators mentioned earlier which determines which discontinuities in images are material changes and not shadow boundaries. This technique is shown to be more accurate and robust than previous methods reported. Another problem addressed is that of color constancy, which is the perceptual ability of the human visual system to assign the same colors to objects under different lighting conditions. We use a method, based on finite-dimensional linear models of reflectance and illumination, which allows the transformation (R,G,B) images into color constant images. In contrast to previous work, we show that good results can be obtained using a 3-receptor system and statistical measurements of natural materials and illuminants. Finally, we address the problem of identifying of highlights in images through the use of chromatic information. We show that understanding reflection, through the use of existing models, allows us to make predictions about the behavior of highlights. In particular, some observations can be made regarding the shift in color from diffuse to specular reflection. Based on these observations, we developed an algorithm which segments images into regions and looks for shifts in color between adjacent regions, and labels the ones which fit the expected relation between diffuse and specular reflection.