Applying image retrieval techniques to large image databases requires the restriction of search space to provide adequate response time. This restriction can be done by means of clustering techniques to partition the image data set into subspaces of similar elements. In this article several clustering methods and validity indices are examined with regard to image categorization. A subset of the COIL-100 image collection is clustered by different agglomerative hierarchical methods as well as the k-Means, PAM and CLARA clustering algorithms. The validity of the resulting clusters is determined by computing the Davies-Bouldin-Index and Calinski-Harabasz-Index. To evaluate the performance of the different combinations of clustering methods and validity indices with regard to semanti-cally meaningful clusters, the results are compared with a given reference grouping by measuring the Rand-Index.
展开▼
