An accurate HMM-based similarity measure between finite sets of histograms

Histogram analysis has nowadays gain in interest, and a lot of work yet address this task. In most of the existing approaches, histograms are manipulated as simple vectors or as statistic distributions. As a consequence, only the bin values of the histograms are mostly considered and the histograms visual shapes are generally neglected. In this paper, hidden Markov models (HMMs) are associated with finite sets of histograms to capture both: the bin values and the visual shapes of the histograms contained in these sets, regardless of their bin sizes. The similarity rate between these HMMs is then used to compare two finite sets of histograms. Experimented in several areas within and beyond machine learning, the proposed approach exhibited relevant performances which outperformed the existing work in the hierarchical classification of the databases GTZAN+ and Corel.