University of Ioannina
Μπεϋζιανή ανόρθωση και ανακατασκευή εικόνων υψηλής ανάλυσης από άγνωστης υποβάθμισης δεδομένα χαμηλής ανάλυσης
Abstract
dc:descriptionThe research topic of this dissertation is the development of stochastic non-stationary image models as image priors used for regularization in restoration and super-resolution problems. The proposed non-stationary image models lead to spatially adaptive regularization, or in other words, non-uniform regularization along the image which depends on the local spatial activity. Furthermore, working in the stochastic framework we employ the Bayesian methodology to solve these inverse problems and in parallel to estimate the model parameters. Thus, development of Bayesian restoration and reconstruction algorithms is another major issue of this dissertation. First, we introduce a new hierarchical (two-level) Gaussian non-stationary image prior. This prior assumes that the residuals of the first order differences of the image, in four different directions, are Gaussian random variables with zero mean and variance that is spatially varying. In this way, the variances manifest the spatial adaptivity mechanism. In order to deal with the resulting over-parameterization of this model, the spatially varying variances are considered random variables (not parameters) and a Gamma hyper-prior is imposed on them, which is conjugate to the Gaussian. To learn this model and infer the image we propose two iterative algorithms. The first is based on the maximum a posteriori estimation (MAP) principle and computes explicitly both the image and the spatially varying variances in all four directions. The second is a Bayesian algorithm that marginalizes the “hidden variables”. Also, the marginalization of the hidden variables produces a Student’s-t distribution. Next, we propose a new Bayesian inference framework for image restoration using a prior in product form. This prior assumes that the outputs of local high-pass filters, (their number is arbitrary), follow again the Student’s-t distribution. Then, a Bayesian inference methodology is proposed that bypasses the difficulty of evaluating the normalization constant of product type priors. The methodology is based on a constrained variational approximation that uses the outputs of all the local high-pass filters to produce an estimate of the original image. In this manner the use of improper priors is avoided and all the parameters of the prior model are estimated from the data. As a next step, we extend the total-variation prior by introducing a new prior which has a number of novel features. More specifically, we introduce a total-variation (TV) prior with spatially varying regularization parameters. In order to avoid the over parameterization, we introduce a Gamma hyperprior for the spatially adaptive regularization parameters of the local TV priors. We also use this prior in a product form, which means that we assume that the outputs of an arbitrary number of high-pass filters are distributed according to this prior. This gives two novel features to the new prior. First, it is explicitly spatially adaptive and thus it is better suited to capture the salient features of the image. Second, it is in product form and has the ability to enforce simultaneously a number of different properties to the image. If the hidden variables of the second layer are marginalized, the resulting density function has a form similar to a Student's-t distribution; thus, we label it as Generalized Student's-t. Due to the complexity of this model, we resort to the variational approximation for Bayesian inference.
Degree
thesis:*- Grantor dc:publisher
- University of Ioannina
- Year dc:date
- 2008
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Χάντας, Ιωάννης
- Chantas, Giannis
Subjects
dc:subject × 18- Ανόρθωση
- Υπερανάλυση
- Ανακατασκευές
- Στατιστική μοντελοποίηση
- Μπευζιανή συμπερασματολογία
- Κανονικοποίηση
- Αντίστροφα προβλήματα
- Restoration
- Super - resolution
- Reconstruction
- Statistical modeling
- Bayesian inference
- Regularization
- Inverse problems
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/17629