Skip to Main content Skip to Navigation

Techniques de réduction de données et analyse d'images multispectrales astronomiques par arbres de Markov

Abstract : The development of astronomical multispectral sensors allows data of a great richness. Nevertheless, the classification of multidimensional images is often limited by Hughes phenomenon: when dimensionality increases the number of parameters of the model grows and the precision of their estimates falls inevitably, therefore the quality of the segmentation dramatically decreases. It is thus imperative to discard redundant information in order to carry out robust segmentation or classification. In this thesis, we have proposed two methods for multispectral image dimensionnality reduction: 1) bands regrouping followed by local projections; 2) radio cubes reduction by a mixture of Gaussians model. We have also proposed joint reduction/segmentation scheme based on the regularization of the mixture of probabilistic principal components analyzers (MPPCA). For the segmentation task, we have used a Bayesian approach based on hierarchical Markov models namely the hidden Markov tree and the pairwise Markov tree. These models allow fast and exact computation of the a posteriori probabilities. For the data driven term, we have used three formulations: 1) the classical multidimensional Gaussian distribution 2) the multidimensional generalized Gaussian distribution formulated using copulas theory 3) the likelihood of the probabilistic PCA model (within the framework of the regularized MPPCA). The major contribution of this work consists in introducing various hierarchical Markov models for multidimensional and multiresolution data segmentation. Their exploitation for data issued from wavelets analysis, adapted to the astronomical context, enabled us to develop new denoising and fusion techniques of multispectral astronomical images. All our algorithms are unsupervised and were validated on synthetic and real images.
Complete list of metadatas
Contributor : Farid Flitti <>
Submitted on : Monday, June 25, 2007 - 4:08:16 AM
Last modification on : Thursday, April 23, 2020 - 2:26:30 PM
Long-term archiving on: : Monday, September 24, 2012 - 10:25:31 AM


  • HAL Id : tel-00156963, version 1



Farid Flitti. Techniques de réduction de données et analyse d'images multispectrales astronomiques par arbres de Markov. Traitement du signal et de l'image [eess.SP]. Université Louis Pasteur - Strasbourg I, 2005. Français. ⟨tel-00156963⟩



Record views


Files downloads