| The objective of modelling in image analysis is to capture the visual characteristics of images in a few parameters so as to understand the nature of the phenomenon generating the images. In this dissertation, novel approaches for modelling the intensity distribution of the gray levels and spatial interaction between the pixels in the observed image will be introduced.; To get an accurate probabilistic model specifying the distribution of gray level in the observed image, a new probabilistic model will be proposed. The proposed probabilistic model is based on modelling the intensity distribution of multi-modal images using a linear combination of discrete Gaussians (LCDGs) with positive and negative components. The Expectation-Maximization (EM) algorithm will be modified to deal with the LCDGs in order to estimate the parameters of the model and a novel EM-based sequential technique will be presented to get a close LCDGs-approximation as an initialization to the modified EM algorithm. Due to both positive and negative components, the LCDGs approximates accurately not only the main body (mode) of empirical density but also its tails. Experiments show that the approach approximates complex empirical densities more accurately than the commonly used probabilistic models with a linear combination that considers positive components only.; Generally, Markov-Gibbs random field (MGRF) models have been success fully used for modelling spatial interactions between various sites of an image. In this dissertation, four novel MGRF models based on calculating the co-occurrences of the observed signals (gray levels) rather than differences will be presented. Moreover, new approaches of accurate identification (estimation of the Gibbs potentials, and the locations of the neighborhood system) for these four models will be introduced. The generic rotation-scaling variant MGRF model is useful for image alignment, whereas the proposed rotation-invariant model is useful for tracking and segmenting moving objects that have small rotational changes from one frame to another. Moreover, this model can be very useful in segmenting objects such as lung nodules, colon polyps, and brain tumors as these objects have a particular appearance model but may appear in different orientation. Furthermore, new analytical parameter estimates for conventional auto-binomial MGRF and joint MGRF of images and region maps are considered for use in solving image segmentation problems. (Abstract shortened by UMI.)... |
