Comparison of Model-Based Segmentation Algorithms for Color Images

摘要

The objective of this thesis is to develop segmentation methods for multichannel and single channel images, and compare these methods. The segmentation algorithms are based on linear model for the image textures and on inverse filtering to estimate the image textures and their regions. Two specific methods are compared 1) A multichannel filtering algorithm that simultaneously models the three separate signals representing the intensity of red, green, and blue as a function of spatial position and 2) A single channel model applied to a combined image resulting from performing a Karhunen-Loeve transformation on the three signal components. Results of the multichannel image segmentation and the Karhunen-Loeve transformed one-channel image segmentation are presented and compared. Keywords: Maximum likelihood; Markov random fields; Computer programs; Theses. (Author)