Stability and inference in discrete diffusion scale-spaces

No Thumbnail Available
Journal Title
Journal ISSN
Volume Title
Faculty of Information and Natural Sciences | Doctoral thesis (monograph)
Checking the digitized thesis and permission for publishing
Instructions for the author
Date
2008-02-29
Major/Subject
Mcode
Degree programme
Language
en
Pages
84
Series
Dissertations in computer and information science. Report D, 25
Abstract
Taking averages of observations is the most basic method to make inferences in the presence of uncertainty. In late 1980's, this simple idea has been extended to the principle of successively average less where the change is faster, and applied to the problem of revealing a signal with jump discontinuities in additive noise. Successive averaging results in a family of signals with progressively decreasing amount of details, which is called the scale-space and further conveniently formalized by viewing it as a solution to a certain diffusion-inspired evolutionary partial differential equation (PDE). Such a model is known as the diffusion scale-space and it possesses two long-standing problems: (i) model analysis which aims at establishing stability and guarantees that averaging does not distort important information, and (ii) model selection, such as identification of the optimal scale (diffusion stopping time) given an initial noisy signal and an incomplete model. This thesis studies both problems in the discrete space and time. Such a setting has been strongly advocated by Lindeberg [1991] and Weickert [1996] among others. The focus of the model analysis part is on necessary and sufficient conditions which guarantee that a discrete diffusion possesses the scale-space property in the sense of sign variation diminishing. Connections with the total variation diminishing and the open problem in a multivariate case are discussed too. Considering the model selection, the thesis unifies two optimal diffusion stopping principles: (i) the time when the Shannon entropy-based Liapunov function of Sporring and Weickert [1999] reaches its steady state, and (ii) the time when the diffusion outcome has the least correlation with the noise estimate, contributed by Mrázek and Navara [2003]. Both ideas are shown to be particular cases of the marginal likelihood inference. Moreover, the suggested formalism provides first principles behind such criteria, and removes a variety of inconsistencies. It is suggested that the outcome of the diffusion should be interpreted as a certain expectation conditioned on the initial signal of observations instead of being treated as a random sample or probabilities. This removes the need to normalize signals in the approach of Sporring and Weickert [1999], and it also better justifies application of the correlation criterion of Mrázek and Navara [2003]. Throughout this work, the emphasis is given on methods that enable to reduce the problem to that of establishing the positivity of a quadratic form. The necessary and sufficient conditions can then be approached via positivity of matrix minors. A supplementary appendix is provided which summarizes a novel method of evaluating matrix minors. Intuitive examples of difficulties with statistical inference conclude the thesis.
Description
Keywords
smoothing, scale-space, discrete nonlinear diffusion, majorization, Rose determinant, doubly stochastic matrix, variation diminishing, total positivity, tridiagonal M-matrix, optimal scale selection, marginal likelihood inference, monotonicity, Gaussian models, Faà di Bruno formula
Other note
Citation
Permanent link to this item
https://urn.fi/urn:nbn:fi:tkk-011270