EDGE-DETECTION BY POINT CLASSIFICATION OF CANNY FILTERED IMAGES

This paper presents a new strategy that exploits a neural classifier t o select candidate edge points from a filtered image. First, a spatial filtering for edge enhancement (the Canny filter) is used to calculat e a set of large variation points, corresponding to the local maxima o f the filtered image. A preliminary coarse selection is then performed , which exploits neighbourhood information to produce an extended pseu do-edges set (PES). Finally, a features' vector is computed for each p oint belonging to the PES, and fed into a classifier that decides whet her it belongs to the target edge set or not. Since the selection work s at the PES level, the creation of data sets for the training and tes ting of the classifier was performed in a fast and easy way by means o f a computer-aided interactive tool. Experimental results proved that the proposed selection criterion is effective in improving the perform ances of the detector over classical threshold methods (e.g., the hyst eresis selection used by Canny). (C) 1997 Elsevier Science B.V.