Analysing and simplifying histograms using scale-trees

Abstract

A new method for analysing image histograms is introduced. The technique decomposes a histogram into probability level sets. The relationships between these level sets are encoded using a tree. The tree has fewer nodes than the histogram and so is a compressed feature. When used in image retrieval experiments the tree is shown to have a performance that is superior to many methods and no worse than the best alternatives. The tree is efficient because it can be built using a computationally efficient algorithm known as a sieve.