Morphological pattern-spectrum classification of noisy shapes: Exterior granulometries

摘要

Moments of morphological granulometric pattern spectra have been used for classification of both texture and patterns. The present paper concerns shape classification in the presence of edge noise when using granulometries generated by linear structuring elements. Although these have been shown to provide excellent texture classification, it is shown here that they cannot successfully discriminate between close convex shapes when there is even modest edge noise. By redefining a linear granulometry so that it operates on the linear convex hull in the direction of the generating structuring element, excellent shape recognition is achieved, even for edge noise exceeding that typically encountered in practice. These newly proposed granulometries are called exterior granulometries and, for a given shape class, optimal feature sets from amongst their first three pattern-spectrum moments are obtained. Their superior performance is demonstrated by a probabilistic comparison between the stability of the pattern-spectrum means for both ordinary and exterior linear granulometries.