Efficient estimation of the number of clusters for high-dimension data

The exponential growth of digital image data has given rise to the need of efficient content management and retrieval tools. Currently, there is a lack of tools for processing the collected unlabeled data in a schematic manner. K-means is one of the most widely used clustering methods and has been applied in a variety of fields, one of them being image sorting. Although a useful tool for image management, the K-means method is heavily influenced by initializations, the most important one being the need to know the number of clusters a priori. A number of different methods have been proposed for identifying the correct number of clusters for K-means, one of them being the variance ratio criterion (VRC). Despite its popularity, the VRC method comes with two very important shortcomings: it only yields good results when the data dimensionality is low and it does not scale well for a high number of clusters, making it very difficult to use in computer vision applications. We propose an extension to the VRC method that works for increased cluster number and high-dimensionality data sets and therefore is fit for image data sets.