Upgrading Color Distributions for Image Retrieval Can We Do Better?

Summary

Content-based image retrieval primarily used color distributions as descriptors of the image content; researches have since focused on the use of various color representation spaces, color and illumination invariance, color quantization and color matching. In order to overcome the many limitations of the description by a firstorder distribution, several higherorder distributions have been introduced since (like autocorrelogram or color coherence vectors). Although they can perform better, their computational complexity is prohibitive and they require parameter setting. We propose to upgrade the first order color distribution (color histogram) by embedding for each color additional information about its perceptual or statistical relevance. Such information is obtained by using local activity measures such as the Laplacian, the entropy and others. We prove that the new color distribution family is compact, robust and easy to compute and provides a superior retrieval performance, independent with respect to the color representation.