An Analysis of Some Models Used in Image Segmentation
Résumé
This report describes an investigation into the characteristics of a number of image models and algorithms used for the segmentation of SPOT images. The initial goals were the problems of phase transition in the model, and parameter estimation. The experimental tools used encompass a great deal of modern Markov chain Monte Carlo (MCMC) methodology, from the Gibbs sampler/Metropolis-Hastings algorithm, to the Swendsen-Wang algorithm and Monte Carlo Maximum Likelihood. One result of this work has been the demonstration of the importance of in-depth studies of the image models being considered -- some common models are shown to be inadequate for the purposes to which they are commonly put. Outlines of future areas of research aimed at overcoming some of the problems identified are given.