Presenter

Antonios Katartzis - ETRO-VUB

Abstract

Bayesian estimation and Markov processes, like Markov chains or Markov random Fields (MRFs), constitute a popular branch in statistics that has been extensively used in the domain of computer vision. The objective of this dissertation is the exploration of new Markovian models for the development of image analysis methods in two important fields: (a) perceptual organization for object identification and (b) pixel- and region-based image classification. The proposed methodologies have been applied on the domains of remote-sensing and medical imaging. They are based on a wide range of image analysis concepts, such as color/multispectral image processing, mathematical morphology, 3-D image geometry, along with the main principles of graph, Gestalt and scale-space theory. The dissertation is organized in three thematic parts.

The First part provides a general introduction to the concepts of Bayesian estimation and Markov processes, emphasizing their potential in solving inverse problems in computer vision. This introductory part includes an overview of the basic Markovian models associated with both 1-D and higher dimensional processes, along with a description of the commonly used methodologies for Bayesian inference and parameter estimation.

The second part highlights the potential of MRF-theory in the case of highlevel vision applications and in particular, those of 2-D and 3-D object detection/recognition. The presented methodologies are based on the principles of perceptual organization. The generation of object hypotheses is formulated as an hierarchical grouping of image primitives, using a set of structural constraints with a successive level of abstraction. These constraints can be efficiently modeled in the MRF framework, giving rise to a wide variety of schemes for shape/structure description and recognition. Based on this strategy, we have developed two methods for the extraction of 2-D and 3-D structural information, using remote sensing imagery. The first method refers to the identification of linear features, like roads and paths but can also be applied to any other image modality. A second application is the detection of 3-D objects corresponding to buildings, using a single remote sensing image and minimal domain knowledge, restricted only to the camera's extrinsic and intrinsic parameters and the time of image acquisition.

Finally, the third part is concerned with low-level vision applications, focusing on the issue of image classification. We present an application dependent, pixel-based classification scheme for the identification of the breast skin thickness in mammographic images and a more general region-based approach for supervised and unsupervised image classification. Both approaches consider a multiscale representation of the image. The first one employs a wavelet-based observation field, in combination with a non-hierarchical MRF prior model that describes the geometrical properties of the breast skin region. The second approach introduces a novel classification framework, in which multiscale information is incorporated in the labeling process in the form of a hierarchical Markovian model that explicitly exploits a multiscale image hierarchy. The novelty of the proposed method is the use of a hierarchical structure, which has the form of a multiscale region adjacency graph. The latter is generated using an image adaptive scale-space filter and a linking scheme for relating image features across the scale dimension.