Classification of visual information by structure based similarity analysis

Two algorithms - for online and for offline classification of images are presented in this paper. They perform the classification by comparing the dissimilarity degrees between all pairs of available images. A special computational way is proposed to evaluate the normalized dissimilarity, based on the color RGB histograms, extracted from each image. The online classification algorithm creates a multi-tree-like structure of the newly and sequentially submitted images that are attached to the preliminary defined classes of pivot (prototype) images. Occasionally they could also create new classes. The offline classification produces rather a general structure of connected subsystems, each of them containing a number of similar mages. Both algorithms are illustrated on a specially created test example of 18 images. The analysis of the results shows the applicability of the algorithms to different practical areas, such as robot vision, visual searches and possibly others.