Abstract

The aim of this report is to establish the credibility of the Gaussian Mixture Model (GMM) as a model for the distributions of oriented edgels of rigid and biological objects in noisy images. This is tackled in two stages: first, the response of the Soble filter to noisy pixels is analysed to show that the result holds for smooth ridid objects. Second, arguments are presented to support the proposition that the model can also effectively capture the added uncertainty introduced by natural shape variation, as found in images of biological objects. The result has particular application in the extension of the Generalized Hough Transform (GHT) to deformable shapes; in particular if offers a tailored and manipulable alternative to the non-parametric kernel density estimate used by Ecabert and Thiran.