Superpixel Clustering for Detection of Binary Object Hierarchy Using Modernized Classical Clustering Methods

Abstract

The paper proposes the simplest non-algorithmic definition of superpixels as image elements, which itself determines the algorithm for their calculation. A system of three classical methods of image approaching by piecewise constant approximations by means of iterative clustering of image pixels is considered: Ward’s clustering, split-and-merge method and Kmeans method. The modernization of these methods is suggested for reduction of the approximation error E (total squared error) to the achievable minimum values for a fixed cluster numbers gin the current approximation. Advanced versions of the classical methods for reducing of the approximation error E are combined in so-called standard model for detecting of binary hierarchy of objects in the image by means of iterative superpixel clustering. In this paper the advanced versions of mentioned methods are presented and the standard model of binary hierarchy of objects in the image is briefly described.