Modèle bayésien non paramétrique pour la segmentation jointe d'un ensemble d'images avec des classes partagées

Abstract : This work concerns the joint segmentation of a set images in a Bayesian framework. The proposed model combines the hierarchical Dirichlet process (HDP) and the Potts random field. Hence, for a set of images, each is divided into homogeneous regions and similar regions between images are grouped into classes. On the one hand, thanks to the HDP, it is not necessary to define a priori the number of regions per image and the number of classes, common or not.On the other hand, the Potts field ensures a spatial consistency. The arising a priori and a posteriori distributions are complex and makes it impossible to compute analytically estimators. A Gibbs algorithm is then proposed to generate samples of the distribution a posteriori. Moreover,a generalized Swendsen-Wang algorithm is developed for a better exploration of the a posteriori distribution. Finally, a sequential Monte Carlo sampler is defined for the estimation of the hyperparameters of the model.These methods have been evaluated on toy examples and natural images. The choice of the best partition is done by minimization of a numbering free criterion. The performance are assessed by metrics well-known in statistics but unused in image segmentation.
Complete list of metadatas

Cited literature [73 references]  Display  Hide  Download

https://tel.archives-ouvertes.fr/tel-01950357
Contributor : Abes Star <>
Submitted on : Monday, December 10, 2018 - 5:24:07 PM
Last modification on : Friday, May 17, 2019 - 11:39:03 AM
Long-term archiving on : Monday, March 11, 2019 - 4:26:06 PM

File

SODJO_JESSICA_2018.pdf
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-01950357, version 1

Citation

Jessica Sodjo. Modèle bayésien non paramétrique pour la segmentation jointe d'un ensemble d'images avec des classes partagées. Traitement du signal et de l'image [eess.SP]. Université de Bordeaux, 2018. Français. ⟨NNT : 2018BORD0152⟩. ⟨tel-01950357⟩

Share

Metrics

Record views

201

Files downloads

64