COMPUTING ILLUMINATION-INVARIANT DESCRIPTORS OF SPATIALLY FILTERED COLOR IMAGE REGIONS

Spatial filters provide a useful and efficient means of analyzing an i nput color image into components that capture different spatial proper ties, Representations based on spatial filtering have restricted usefu lness for recognition, however, because the output of a spatial filter across an Image depends on the scene illumination conditions, In this paper, we use a physically accurate linear model for spectral reflect ance to derive invariants of distributions in spatially filtered color images that do not depend on the scene illumination. These invariants can be used for the illumination-invariant recognition of regions fol lowing an arbitrary linear filtering operation, We describe a method f or illumination correction based on color distributions and introduce an illumination change consistency constraint that is useful for verif ying matches obtained using the invariants. We show, using a set of cl assification experiments, that the filtered distribution invariants ca n significantly improve the capability of a recognition system in envi ronments where illumination cannot be controlled.