The state of color vision research

Abstract

The members of the Computational Vision Laboratory at Simon Fraser University have been studying color for over a decade. I will discuss some of the main color issues, the progress we have made in understanding them and the application of our methods to color-based object recognition and digital photography.

To explain color perception, we must explain how it is that we see colors as relatively stable despite changes in the incident illumination. I make the assumption that color, like the rest of visual perception, is there to give us information about the world, the surface properties of objects in particular, and so the stability and reliability of the information is important. The problem of color stability arises because the light reaching our eyes from an object is the product of the object's surface reflectance and the spectrum of the light illuminating the object. We do not have direct access to the properties of the incident light, so somehow we must estimate them from the light we receive from the object. To make matters worse, our eyes only measure the spectrum at extremely low resolution. Clearly a stable representation of object color would be useful in computer vision. It is equally important for digital photography since we do not want images to produce images whose color balance depends on the qualities of the ambient scene illumination. Until recently the growth of digital photography was limited less by camera technology than by the lack of economical digital printers with photographic quality, a situation that has changed dramatically in the last six months with the introduction of some new ink jet printer technology (e.g. the Hewlett-Packard Photo Smart printer). Digital photography is providing new impetus to the search for better models of color perception.

Comparative tests of many different color constancy methods show that two work most reliably: one based on a neural network to estimate the illumination properties from a color histogram, and a second based on the constraints provided by the image gamut.