2000
@phdthesis{Zac2000,
vgclass = {thesis},
vgproject = {cbir},
author = {John M. Zachary},
title = {An Information Theoretic Approach to Content Based Image Retrieval},
school = {Department of Computer Science, Louisiana State
University},
address = {298 Coates Hall, Baton Rouge, LA 70803, USA},
month = {September},
year = {2000},
url = {http://citeseer.nj.nec.com/zachary00information.html},
abstract = {We propose an information theoretic approach to the
representation and comparison of color features in digital images to
handle various problems in the area of contentbased image retrieval.
The interpretation of color histograms as joint probability density
functions enables the use of a wide range of concepts from information
theory to be considered in the extraction of color features from images
and the computation of similarity between pairs of images. The entropy
of an image is a measure of the randomness of the color distribution in
an image. Rather than replacing color histograms as an image
representation, we demonstrate that image entropy can be used to
augment color histograms for more efficient image retrieval. We propose
an indexing algorithm in which image entropy is used to drastically
reduce the search space for color histogram computations. Our
experimental tests applied to an image database with 10,000 images
suggest that the image entropy-based indexing algorithm is scalable for
image retrieval of large image databases. We also proposed a new
similarity measure called the maximum relative entropy measure for
comparing image feature vectors that represent probability density
functions. This measure is an improvement of the Kullback-Leibler
number in that it is non-negative and satisfies the identity and
symmetry axioms. We also propose a new usability paradigm called Query
By Example Sets (QBES) that allows users, particularly novice users,
the ability to express queries in terms of multiple images.},
}