Statistical mechanical models for image processing

This thesis introduces a solution to the problem of image restoration and feature extraction by incorporating new image models derived from statistical physics. Starting from Shannon's model of information processing, a special lattice spin Hamiltonian is used which is well suited for both source coding and for modeling information loss within the Bayesian framework. By applying a high-temperature expansion the parameter estimation problem is solved analytically using transfer-matrix methods. A Monte Carlo simulation restores the distorted image utilizing the statistical information about the source and the channel. Beyond its inherent practical usefulness the image restoration problem illustrates directly basic concepts related to information theory, statistical inference, and perception. The work is split mainly in two parts: Chapters 1 to 4 contain a summary of the problem and existing models, Chapters 5 to 7 introduce the new models and illustrate their capabilities in a variety of experiments. The conclusion can be found in Chapter 8.

Level thesis:degree_level

thesis.doctoral

Grantor dc:publisher

Universität Oldenburg

Year

2001

Author dc:creator

• Wanschura, Thorsten

• [Keine Schlagwörter von Autor/in vergeben.]

Repository record source_url

http://oops.uni-oldenburg.de/353

OAI identifier oai:identifier

oai:oops.uni-oldenburg.de:353