Bayesian scale space analysis of images

Abstract

Two new statistical scale space methodologies are discussed. The first method aims to detect differences between two images obtained from the same object at two different instants of time. Both small scale sharp changes and large scale average changes are detected. The second method detects features that differ in intensity from their surroundings and it produces a multiresolution analysis of an image as a sum of scale-dependent components. As images are usually noisy, Bayesian inference is used to separate real differences and features from artefacts caused by random noise. The use of the Bayesian paradigm facilitates application of flexible image models and it also allows one to take advantage of an expert's prior knowledge about the images considered.