Color constancy using KL-divergence

Color is a useful feature for machine vision tasks. However its effectiveness is often limited by the fact that the measured pixel values in a scene are influenced by both object surface reflectance properties and incident illumination. Color constancy algorithms attempt to compute color features which are invariant of the incident illumination by estimating the parameters of the global scene illumination and factoring out its effect. A number of recently developed algorithms utilize statistical methods to estimate the maximum likelihood values of the illumination parameters. This paper details the use of KL-divergence as a means of selecting estimated illumination parameter values. We provide experimental results demonstrating the usefulness of the KL-divergence technique for accurately estimating the global illumination parameters of real world images.

1 Introduction

In machine vision and image database applications color can be used as a simple means of segmenting or identifying a specific object as described in [14] or as a means of quickly identifying likely candidate regions for object recognition as in [18]. However, problems arise when images are captured under varying illumination conditions and with cameras with differing sensor characteristics. Color constancy methods try to overcome these problems by estimating the surface reflectance properties of objects in a scene regardless of scene illumination and camera characteristics. In general, if we assume diffuse surface reflections, the measured pixels values for camera sensor channel at image location , denoted as , are the product of the incident illumination , the surface reflectance properties and camera sensor channel spectral sensitivity as a function of the wavelength of the incident light , integrated over the visible spectrum, :

$$\rho_{k}(i)={\int}_{\omega}E(\lambda)S(i,\lambda)C_{k}(\lambda)d\lambda$$