Image Processing for Space & Earth Observations
Multisource & Multispectral Image Processing via Bayesian Inference
MIV team ICUBE

CNRSUnistra

The IPSEO Research Group

IPSEO aims at developing new data analysis tools for remote sensing and astronomical imaging domain. Probabilistic approaches are used to model complex datasets (multispectral, multiple sources), and to solve the inverse problems related to their analysis.

IPSEO is currently a research group within the MIV team (Modèle Image Vision) of the Laboratoire des Sciences de l'Ingénieur, l'informatique et de l'Image, ICUBE (Illkirch)(former LSIIT Laboratory). It collaborates with the Centre de Données Astronomiques de Strasbourg (CDS,Strasbourg Observatory), mainly funded by CNRS, Université de Strasbourg and various national and European projects. It is binding new collaborative links with the SERTIT team, a new partner inside ICUBE Laboratory.

  • Research Goals & Objectives
  • Projects: Applications / Data type
  • Our Probabilistic Approach
  • Highlights – Recent Results
  • Research Goals & Objectives

    The main goal is to develop new image analysis methods to answer specific needs for remote sensing and space sciences specialists, e.g. the hyperspectral and/or multimodal image analysis in astronomy and earth sciences. The increasing wealth, complexity and size of imaging data in these fields are major factors in addition to scientific interpretation that we take into account in our developements.

    Data understanding • Accurate inference • Efficient analysis tools

    The Aladin Sky Atlas (CDS) is a Virtual Observatory interface, integrating many visualisation and processing tools.
    • The data are modeled within a probabilistic framework, thus enabling us to integrate multiple sources and modalities, possibly multi- or hyperspectral depending on the sensor type. [more details]
    • Bayesian inference is used to estimate the quantities of interest as well as their uncertainties, thus allowing for accurate data analysis, model checking and assessment. [more details]
    • We seek to develop efficient algorithms and fast implementations, in contact with specialists of each area. New tools for astronomical data are checked by astronomers collaborating to our projects and integrated when applicable into a Virtual Observatory framework through collaborations with the CDS and CRAL. Remote sensing applications take place within our SERTIT partnership but are also seeked for towards other research institutions or industrial partners.

    Research Projects: Application areas / Data type

    Our research projects are summarized in the following table. Projects are displayed according to the target application domain (columns) and source image type to be processed (rows). [click buttons for details]

    APPLICATION Astronomy,
    Astrophysics
    Planetary Imaging
    (planets, small bodies)
    Remote Sensing
    (Earth Observation)
    DATA TYPE
    image
    sources
    spectral
    depth
    single
    image
    mono
    band
    PNSpectrum
    SurfaceModelRender
    SplitSplatRender
    MTFInferFractal
    DispMapInference
    BayesCameraCalibra...
    multiple
    images
    Astro Metadata
    3DShapeInference
    single
    image
    multi
    band
    QTreeMultiSegment
    MissingMultiSegmentA
    FuzzyMarkovSegment
    QTreeRestoFusion
    CopulaHMTSegment
    LSBGalaxyDetect
    GalaxyClassification
    SpectroDec
    MissingMultiSegmentE
    CorrNoiseWaterSegm...
    ImPolar
    multiple
    images
    DeepSkyFusion
    3DSpaceFusion
    ReflectanceFusion
    single
    image
    hyper
    spectral
    HyperGalaxyClass
    MultiColorViz
    AstroCubeSegmentat...
    CubeDeconvolution
    CubeSourceSeparation
    SpatioSpectralSEG
    DSIM
    multiple
    images
    CubeCombinaisonFus...
    CubeVisualization
    CubeExtract
    CubeExtract
    AstroCubeLineSeg
    DetectChangeUrbanM...
    LowSNRSpatSpecDetect

    An Interdisciplinary Research

    Our application areas range from deep sky exploration to Earth observation, aiming at a better understanding of space imaging in order to provide new tools for the specialists of these areas.

    • Astronomy (star fields, galaxies, nebulae) → image enhancement and improved visualisation (e.g. denoising, color coding), data fusion for an easier data manipulation (e.g. channel fusion, multi-source fusion, data reduction), automatic analysis and physics-based interpretation (e.g. segmentation, classification).
    • Planetary Imaging (Solar System planets, satellites, small bodies) → image enhancement (e.g. deblurring, PSF estimation), model-based data fusion (e.g. 3D surface reconstruction, super-resolution), reflectance estimation and analysis (e.g. terrain classification).
    • Remote Sensing (Earth observations at various resolutions) → shadow and vegetation segmentation in urban landscape; (former) image enhancement (e.g. blind deconvolution), model-based data fusion (e.g. topography reconstruction, super-resolution, data reduction), bidirectional reflectance estimation and analysis (e.g. segmentation).

    Analyzing Various Data Types & Modalities

    Depending on the application, various types of input images are used and the processing methods must adapt consequently. We separate the methods into three main classes based on the spectral depth (one, several, many bands). For each class we have two subclasses, whether a single image or multiple ones are required.

    • Image sources: (single image vs. multiple images) → One input image is used in classical techniques (e.g. deblurring, segmentation). However some tasks require at least two images (e.g. surface recovery), other require even more (e.g. bidirectional reflectance inference). In some cases the task itself is dictated by the great number of available sources (e.g. data fusion). Multiple modalities, other than image-like data, can also be accounted for (e.g. existing elevation models, radar measurements).
    • Spectral depth: (single band vs. multi or hyperspectral) → Single band images are the simplest to handle, whereas multispectral images require to account for band interactions and increased dimensionality of the problem (Hughes phenomenon). In addition, hyperspectral data usually require dimensionality reduction prior to further processing and analysis.

    Our Probabilistic Approach

    Probability theory provides a rigorous setting for expressing the uncertain nature of the objects of interest (e.g. images, surfaces, parameters), due to the randomness of the data acquisition process, as well as epistemic uncertainties in the modeling. In this context, all quantities are random variables, whose probability density function (pdf) is sought, often through approximations and statistics computed from the data.

    Graphical models are among the tools we use to model the image formation process.
    • Using Appropriate Image Models

      Our approach is model-based, since it enables us to include all the available knowledge into a formal, a priori model (e.g. fractal priors for natural images using wavelets, Markov random fields). Furthermore, understanding the image formation process (from the 'perfect' image model to the recorded data), is essential to any analysis task. Therefore, we first need to build a physically meaningful forward model, leading from the analysis objective (e.g. class labels, 3D surface) to the possibly corrupted, noisy observation(s).

    • The Bayesian Inference Approach to Inverse Problems

      Bayesian inference consists of finding the pdf for all parameters of interest, knowing all the observations — the posterior pdf. The mean gives the optimal parameters; the spread relates to the uncertainties. Posterior computation often requires marginalization, i.e. integration w.r.t. all other variables. This is non-trivial and may need efficient approximation techniques.
      We consider image analysis as the formal inversion of a forward model, which is usually an ill-posed inverse problem. Thus, Bayesian inference is an elegant way to solve such a problem, since a priori models help constrain and stabilize it.

    • Designing Efficient Algorithms

      We aim at the computationally efficient analysis of space imaging data. This implies the design of deterministic, efficient algorithms, although Monte-Carlo techniques may be used for research and validation purposes. Our topics of interest therefore include optimization methods (e.g. multigrid, hierarchical, genetic) and approximation theory. Validated implementations are to be integrated into web interfaces (e.g. Aladin @ CDS, Marsiaa project), to be freely used by the research community.


    Webmaster (C. Collet), (M.Louys)
    © IPSEO Group 2005-2016 | WebSite Info & Credits
    Last update: Dec 8, 2016
    IPSEO
    Home
    Research projects
    Highlights-Results
    Softwares

    People
    Collaborations

    Local Admin