Image clustering and retrieval combining fixed/adaptive-binned histograms and various distance functions

In the context of content-based image retrieval, we compare two types of histograms, fixed and adaptive, both frequently used for modeling the image features. We demonstrate that a choice of a histogram type, combined with the choice of a distance function, can have a huge impact onto the clustering structure of the dataset. Such a hierarchical clustering structure visualization of database objects helps often the user to find similar objects and discover unknown patterns. In our experiments we use real data sets with large number of semantic categories, and evaluate both the reachability plots and the clustering accuracy, to show the effects of appropriate choice of fixed and/or adaptive binning in combination with various distance functions. Results show that significant clusters, along with their representatives, can be automatically extracted, which is a basis for visual data mining but even more important for nonvisual data mining.