Skip to Main content Skip to Navigation

Proposition d'une approche de segmentation d'images hyperspectrales

Abstract : Hyper-Spectral Imaging (HIS) is the acquisition and the analysis of images of the same view for a set of wavelengths, in a given spectral domain. Such images raise some information on the chemical constitution of objects thus allowing to differentiate objects of same colour but with different chemical composition. Yet, whatever is the application domain, most of the methods developed for HIS processing, perform data analysis without taking into account the spatial information. Each pixel is considered individually, as a simple array of spectral measurements, with no specific arrangement. Yet, considering the available spectral and spatial information altogether, as a consequence of new spectral image analysers with a higher resolution, appears to be of a main importance to allow the processing of complex images.
Inside our HIS work field, we got involved with such a problematic. And thus, we proposed an iterative schema called butterfly, allowing to tight closely the spectral and spatial aspects in a symmetric way, so not to enforce one space against another and in a conjunctive way as to enable the spatial and spectral information to feed each other in the task of image processing. On this process, the spectral and spatial combination proposed resides in one hand, in building a spatial space adapted to the topological information (latent variables) and in another hand, in a topology generation on the spatial space. In our recent developments, we limit the topological aspect to the region concept.
The proposed approach appears to be a generic proceeding, where one can choose the suitable content and implementation to the processed images. In the subject field of this thesis, we proposed an implementation using a two step process, consisting of a division followed by a fusion. The building of an adjusted space for both steps was realised respectively by diagonalising the intra-region and inter-region matrices. For the division step, we also tested many segmentation approaches, among which the spatially constrained normalised cuts and a modified version of the watershed algorithm. At the end, as to show how it is operating and its main characteristics, we processed our approach with some synthetic as well as with some real images.
Document type :
Complete list of metadatas
Contributor : Nathalie Gorretta <>
Submitted on : Monday, March 16, 2009 - 12:04:31 PM
Last modification on : Friday, July 3, 2020 - 3:39:16 AM
Long-term archiving on: : Tuesday, June 8, 2010 - 11:28:58 PM


  • HAL Id : tel-00368348, version 1


Nathalie Gorretta. Proposition d'une approche de segmentation d'images hyperspectrales. Interface homme-machine [cs.HC]. Université Montpellier II - Sciences et Techniques du Languedoc, 2009. Français. ⟨tel-00368348⟩



Record views


Files downloads