A clustering method based on multidimensional texture analysis

Considering the analogy between image segmentation and cluster analysis, the aim of this paper is to adapt statistical texture measures to describe the spatial distribution of multidimensional observations. The main idea is to consider the cluster cores as domains characterized by their specific textures in the data space. The distribution of the data points is first described as a multidimensional histogram defined on a multidimensional regular array of sampling points. In order to evaluate locally a multidimensional texture, a co-occurrence matrix is introduced, which characterizes the local distribution of the data points in the multidimensional data space. Several local texture features can be computed from this co-occurrence matrix, which accumulates spatial and statistical information on the data distribution in the neighborhoods of the sampling points. Texture features are selected according to their ability to discriminate different distributions of data points. The sampling points where the local underlying texture is evaluated are categorized into different texture classes. The points assigned to these classes tend to form connected components in the data space, which are considered as the cores of the clusters.