Skip to Main content Skip to Navigation

Filtrage, réduction de dimension, classification et segmentation morphologique hyperspectrale

Abstract : Hyperspectral image processing is the generalization of analysis of color images, having three components red, green and blue, to multivariate images having several tens or hundreds of components. In a general way, hyperspectral images are not only acquired in the domain of wavelengths but correspond to a pixel description by a set of values: i.e. a vector. Each one of the components of an hyperspectral image is a spectral channel, and the vector which is associated to each pixel is called spectrum. In order to validate the generality of our processing methods, we have applied them to several kinds of imageries corresponding to the most various hyperspectral images : some pictures with a few tens of components acquired in the domain of wavelengths, some satellite images of remote sensing, some temporal series of Dynamic Contrast Enhanced Magnetic Resonance Imagery (DCE-MRI) and some temporal series of thermal imagery. During this PhD, we have developed a complete chain for automatic segmentation of hyperspectral images by morphological technics. In order to realize it, we have perfected an efficient method of spectral denoising using Factor Correspondence Analysis (FCA). This method preserves spatial contours, which is very useful for morphological segmentation. Then, we have reduced the dimension of the image using data analysis methods, or using spectral modelling, in order to get another representation of the image with a reduced number of channels. Starting from this image of smaller dimension, we have made a classification (supervised or not) in order to group pixels into homogeneous spectral classes. However, the obtained classes being not spatially homogeneous, i.e. connected, a segmentation stage is necessary. We have demonstrated that the recent method of the Probabilistic Watershed is particularly adapted to the segmentation of hyperspectral images. It is based on different realizations of markers, conditioned by the spectral classification, to get some realizations of contours by Watershed. These contours realizations ensure the estimate of a probability density function (pdf) which is very easy to segment by a standard Watershed. Finally, the probabilistic Watershed is conditioned by the spectral classification. Therefore, it produces spatio-spectral segmentations with very smooth contours. This treatment chain has been applied on series of DCE-MRI and gives us the possibility to build an automatic method for computer aided detection of cancerous tumors. Moreover, some other segmentation approaches have been developed for hyperspectral images: η-bounded regions and µ-geodesic balls. Thanks to the introduction of regional information, they improve the segmentations by quasi-flat zones which are only based on local information. Finally, we have perfected a very efficient method to compute all the pixels pairs of geodesic distances in an image. It reduces until 50 % the number of operations compared to a naïve approach and until 30 % compared to other methods. The efficient computation of this table of distances opens very promising perspectives for spatio-spectral dimensionality reduction.
Document type :
Complete list of metadatas

Cited literature [158 references]  Display  Hide  Download
Contributor : Ecole Mines Paristech <>
Submitted on : Monday, December 15, 2008 - 8:00:00 AM
Last modification on : Wednesday, October 14, 2020 - 3:52:22 AM
Long-term archiving on: : Saturday, November 26, 2016 - 6:18:36 PM


  • HAL Id : pastel-00004473, version 1


Guillaume Noyel. Filtrage, réduction de dimension, classification et segmentation morphologique hyperspectrale. Mathématiques [math]. École Nationale Supérieure des Mines de Paris, 2008. Français. ⟨NNT : 2008ENMP1558⟩. ⟨pastel-00004473⟩



Record views


Files downloads