Effectiveness of Content-Based Image Clustering Algorithms

To be able to retrieve a set of image documents, the images must be clustered according to semantic similarity. The image clusters are utilized by content-based image retrieval and querying systems that require effective query matching in large image databases. This book describes the clustering process, focusing on clustering algorithms and ways to measure the quality of the created clusters. Four common image clustering algorithms are evaluated: k-means clustering and three versions of hierarchical clustering, using average-linkage, complete-linkage, and Ward''s method. In the experimental section of the book, the algorithms are compared using two similarity measures: color-based similarity utilizing MPEG-7 color descriptors only, and total similarity as a weighted sum of features for both color, texture and shape. The experiments show average-linkage hierarchical clustering performing best according to both similarity measures. Notably, though, the addition of texture and shape features degraded the cluster quality of all the three hierarchical methods tested, but improved the quality of k-means clustering.