Distributional properties and estimation in spatial image clustering (Q2008613)

In machine learning and pattern recognition theory the conventional clustering analysis is point-wise and deals with discrete data points by looking for their associations with specific but \textit{a priori} unknown groups. This papers considers cluster analysis from the perspective of image segmentation problem where spatial relations are important. The distribution of the size of an individual cluster is defined without defining a random variable to be the size of an individual cluster. The main results of this paper are the new methods to estimate the spatial distribution function because without a distribution function of the clusters, it is hard to deal the statistical properties of the data of image clustering and perform efficient statistical inferences. The theoretical results are illustrated by experiments on the real word dataset related to spatial patterning among savanna trees.