Reiner Lenz
Dept. Electrical Engineering
Linköping University
S-58183 Linköping, Sweden
Peter Meer
Dept. Electrical and Computer Engineering
Rutgers University
Piscataway, NJ 08855-0909, USA
Color is one of the main attributes used for content based image retrieval. Similar scenes may not be recovered when captured under different conditions, since the recorded color depends on the conditions under which the image was taken. We have developed a statistical technique which compares two images and estimates their difference under the form of a relative illuminant. The technique does not require pixelwise registration. The first few principal components of the log-spectra of a color appearance system define a space in which the effect of illumination can be approximated by a constant shift. The components of the shift vector are estimated using a robust mode detector, and define the spectrum of the relative illuminant. The relative illuminant can be used either as a global representation of a color image, or for the normalization of a pair of images.
If you want a postscript version of the paper press here (3.5MB!)