Grouping Edgels into Structural Entities Using Circular Symmetry, the Distributed Hough Transform, and Probabilistic Non-accidentalness

Abstract

Grouping disconnected edgels and forming quantitative descriptions of structural entities are fundamental tasks in image understanding. The descriptions sought are based on the assumption that boundaries of man-made objects can be represented by circular arcs and straight lines. We have developed a new grouping algorithm called the distributed Hough transform (DHT). The DHT is capable of grouping edgels into circular features and generating explicit quantitative descriptions of these features. The principle of proximity weighted circular symmetry (PWCS) is incorporated into the DHT, and it is based upon non-accidentalness, viewpoint invariance, and recent probabilistic models of projected angles and distances. The DHT requires a one-dimensional parameter space to detect circular arcs, whereas the conventional Hough transform requires a three-dimensional parameter space. The DHT also has the advantage of being able to consider the important property of proximity; the conventional Hough transform lacks this ability. The DHT is an inherently parallel process and this facilitates its implementation in a neural network. Experimental results on synthetic and true images show that the DHT is a robust algorithm.