Statistical shape and appearance models for segmentation and classification

In this dissertation we develop and apply models of shape and models of image intensities (appearance models) in object-based image processing tasks. We make contributions in three areas of interest: constructing novel flexible models of shape and of image intensities, using these models to extract object boundaries from images, and analyzing differences between groups of shapes from given, extracted object boundaries.;In the shape and appearance model construction and application areas of focus we are motivated by the task of extracting the object boundaries from images by an evolving closed curve technique named curve-evolution. We develop and apply novel models of shape and models of appearance for incorporation in such curve-evolution-based object boundary extraction. In our first major contribution, we start with the statistical shape model based on maximum entropy principle and designed to capture perceptual shape similarity of training shape samples. In sampling experiments, this statistical shape model has been shown to generate new shape samples with prominent visual features of the original training shapes used to construct the model. For the first time, we develop methods to incorporate this maximum entropy model into object boundary extraction tasks. We show that indeed incorporation of such a prior can have a dramatic effect in object boundary extraction problems, favoring the solution similar to the training shapes.;In our next major contribution, we develop a new model of shape based on the notion of shape distributions. Shape distributions have been introduced as cumulative distribution functions of parameters continuously defined on contours. Shape distributions have been used before for shape classification tasks, but our work is their first use for object boundary extraction. The resulting shape models show an excellent ability to preserve prominent visual object structures during boundary extraction in challenging segmentation problems involving high noise, object obstruction, and weak or even missing intensity edges. Further, these models exhibit robustness to limited training data. These models eliminate the need for shape alignment at the model construction and estimation steps, often a difficult and critical task. We further extend these models to capture information on the relative configurations of multiple contours, which helps to extract multiple boundaries more efficiently. We also extend the shape distribution concept to model image intensities. (Abstract shortened by UMI.).