PUB - Publikationen an der Universität Bielefeld

Large image collections require efficient organization and visualization. This paper describes an approach to establish image categories automatically by unsupervised learning. The method works free of context and previous knowledge: in a first stage, features are formed automatically, then images are clustered to form categories. The human database designer has to decide only whether a category is useful or too inhomogeneous from a high level point of view. To collect images that cannot be categorized automatically, an additional ‘miscellaneous’ category exists. Categories are visualized by displaying the most typical image(s) of the categories as thumbnails. The main benefit of the approach is that it deals with color and shape in a unified way on a local scale, combined with the advantages of histogram techniques on the global scale. To judge results, an evaluation scheme which is adequate for the task of categorization is proposed.