An algorithm for pattern description on the level of relative proximity

摘要

This paper presents a purely data-driven and knowledge-free cluster algorithm which formalizes the Gestalt rule of proximity as it is encountered in dot figures. The concept of neighbourhood with respect to relative proximity is described by a binormal distribution function. This spread function expresses the strength with which a point operates upon its surroundings. Summing all partial functions yields a primary description of the dot figure which can be viewed as a landscape with mountains where dots are relatively close to one another, and valleys where they are relatively far apart from each other. Clusters, boundaries and shapes emerge in regions where the resultant function surpasses a threshold value. The proposed method is discussed on the basis of a few examples. The method is further discussed in relation to smoothing techniques for the visual enhancement of noisy patterns and to propagation models for the description of visual form.