Skip to Main content Skip to Navigation
Theses

Approche bayésienne en séparation de sources. Applications en imagerie

Abstract : My thesis consists in applying the Bayesian approach to the source separation problem. My contributions have both theoretic and algorithmic aspects that are illustrated by applications in satellite imaging and in astrophysics. - Theoretic contributions: 1. We have proposed a pertinent source model. The hierarchical aspect of this model is well adapted to the natural hidden structure of the source separation problem. 2. We have characterized the degeneracy of the maximum likelihood in the multivariate and the source separation context. 3. We have proposed an original contribution for the prior selection with the differential geometry tools. - Algorithmic contributions: 1. We have proposed algorithms for the simultaneous separation and segmentation of sources. The common principle of this algorithms is the exploitation of the non stationarity in the temporal, spatial, spectral or time-frequency domains. 2. We have implemented a parallel Gibbs sampling and other stochastic approximations of the EM algorithm. 3. The performances of our method is illustrated with applications in satellite imaging and in astrophysics for the estimation of the CMB (Cosmic Microwave Background). Finally, I open some theoretic perspectives on the duality between the Bayesian approach and the information theoretic approach in the context of the simultaneous separation and segmentation of sources.
Complete list of metadatas

https://tel.archives-ouvertes.fr/tel-00009634
Contributor : Hichem Snoussi <>
Submitted on : Friday, July 1, 2005 - 12:10:06 PM
Last modification on : Wednesday, October 14, 2020 - 3:56:41 AM
Long-term archiving on: : Friday, April 2, 2010 - 10:24:35 PM

Identifiers

  • HAL Id : tel-00009634, version 1

Collections

Citation

Hichem Snoussi. Approche bayésienne en séparation de sources. Applications en imagerie. Traitement du signal et de l'image [eess.SP]. Université Paris Sud - Paris XI, 2003. Français. ⟨tel-00009634⟩

Share

Metrics

Record views

306

Files downloads

611