Abstract
A data-driven multi-dimensional histogram method is proposed for segmentation of images observed in a multi-dimensional space. An image is observed as a mixture of multi-variate data, but the number of clusters in the mixture is generally unknown. In addition, the histogram width needs to be determined by a criterion on the basis of the observed image data information. The proposed algorithm shows criteria to determine the histogram width and the number of clusters according to the observed multi-variate data. Simultaneously, the statistics for each cluster are computed, which are used for clustering all of the other elements. The proposed clustering method can be applied to various kinds of images, because it is not restricted by an image model used in model-based approaches. The effectiveness of the algorithm is evaluated by computer simulation, and the algorithm is aplied to real color image clustering.
