On the use of neural networks and geometrical criteria for localisation ofhighly irregular elliptical shapes

Detection of elliptical shapes is of extreme importance in several computer vision applications. In this paper a new method fur irregular elliptical s hapes localisation in multi-connected regions is described. This method fir st computes a set of elementary are segments, which is then aggregated usin g geometrical decision criteria and a posteriori aggregation probabilities obtained from a neural network for Bayes classification. To identify and ch aracterise the elementary are segments, a cluster identification, a contour grouping strategy and some extensions to Fitzgibbon's ellipse fitting meth od are introduced. These methods are applied successfully in the set-up of an automatic lime granule inspection system. The algorithm has proven to be very robust, since it is able to correctly detect elliptical shapes even w hen noisy data are present.