---------
Reference
---------
Charles Rosenberg, "Image Color Constancy Using EM and Cached
Statistics", to appear in ICML '00.
--------
Abstract
--------
Cached statistics are a means of extending the reach of traditional
statistical machine learning algorithms into application areas where
computational complexity is a limiting factor. Recent work has shown
that cached statistics greatly reduce the computational requirements
of building a mixture model of a distribution using
Expectation-Maximization for a small trade off in model error. This
paper describes a method whereby a mixture model built using cached
statistics is used as a means of improving the color normalization
performance of two standard color constancy algorithms. Color
constancy algorithms factor out illumination effects such that
normalized pixel color values become an invariant representation of
surface reflectance properties. This can improve the performance of
machine vision and image database algorithms which use color as a
feature. This processing is also important in digital camera and
scanning applications where a preferred rendition of a scene is to be
realized independently of the lighting conditions at the time of image
capture. The details and experimental evaluation of two modified color
constancy algorithms which utilize a parametric mixture model are
described.