Robust color image matching based on the norm in the quotient space

摘要

Hue information is effective in identifying observed color images; after extracted from the color images, hue data should be normalized with due consideration for the circularity inherent in the hue information. To perform the normalization optinally, we represent a class of hue as an element in a quotient space, where a norm of the element is defined suitably for human color perception. Although we can also define a dissimilarity measure between sets of the data naturally in the space, there is a problem that the dissimilarity measure is influenced by the normalized datum whose hue angle is around its boundary in the presence of noise. To overcome the problem, we propose a new dissimilarity measure designed via the robust estimation. Experimental results sustain the robustness of the proposed dissimilarity measure.