Skip to Main content Skip to Navigation

Morphology, Geometry and Statistics in non-standard imaging

Abstract : Digital image processing has followed the evolution of electronic and computer science. It is now current to deal with images valued not in {0,1} or in gray-scale, but in manifolds or probability distributions. This is for instance the case for color images or in diffusion tensor imaging (DTI). Each kind of images has its own algebraic, topological and geometric properties. Thus, existing image processing techniques have to be adapted when applied to new imaging modalities. When dealing with new kind of value spaces, former operators can rarely be used as they are. Even if the underlying notion has still a meaning, a work must be carried out in order to express it in the new context.The thesis is composed of two independent parts. The first one, "Mathematical morphology on non-standard images", concerns the extension of mathematical morphology to specific cases where the value space of the image does not have a canonical order structure. Chapter 2 formalizes and demonstrates the irregularity issue of total orders in metric spaces. The main results states that for any total order in a multidimensional vector space, there are images for which the morphological dilations and erosions are irregular and inconsistent. Chapter 3 is an attempt to generalize morphology to images valued in a set of unordered labels.The second part "Probability density estimation on Riemannian spaces" concerns the adaptation of standard density estimation techniques to specific Riemannian manifolds. Chapter 5 is a work on color image histograms under perceptual metrics. The main idea of this chapter consists in computing histograms using local Euclidean approximations of the perceptual metric, and not a global Euclidean approximation as in standard perceptual color spaces. Chapter 6 addresses the problem of non parametric density estimation when data lay in spaces of Gaussian laws. Different techniques are studied, an expression of kernels is provided for the Wasserstein metric.
Document type :
Complete list of metadata

Cited literature [98 references]  Display  Hide  Download
Contributor : ABES STAR :  Contact
Submitted on : Thursday, June 23, 2016 - 6:18:28 PM
Last modification on : Wednesday, November 17, 2021 - 12:27:13 PM
Long-term archiving on: : Saturday, September 24, 2016 - 11:32:09 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01336796, version 1


Emmanuel Chevallier. Morphology, Geometry and Statistics in non-standard imaging. Optimization and Control [math.OC]. Ecole Nationale Supérieure des Mines de Paris, 2015. English. ⟨NNT : 2015ENMP0082⟩. ⟨tel-01336796⟩



Record views


Files downloads