
Séminaires
Les séminaires du projet Ariana ont lieu à
l'INRIA Sophia Antipolis (plan),
la salle ainsi que les résumés (en français
ou/et en anglais) étant affichés dès que possible.
Si vous le souhaitez, vous pouvez consulter l'agenda des séminaires
des années précédentes :
2018, 2017, 2016, 2015, 2014, 2013, 2012, 2011, 2010, 2009, 2008, 2007, 2006, 2005, 2004, 2003, 2002, 2001, 2000, 1999, et 1998. Anciens séminaires du projet Ariana :
Titre 
Intervenant 
Date/Lieu 
Résumé 
Dualtree Mband wavelet analysis; applications to image denoising 
Caroline Chaux postdoc Ariana University of Marne La Vallée 
18/12/2006 14h30 Coriolis 

Résumé (anglais) :
We propose here new waveletbased geometric transforms: the dualtree Mband wavelet analyses. These decompositions allow to perform a multiscale, directional and local analysis of images. They are in the trend of recent works aiming at better representing geometric informations (textures, contours) and to perserve it while processing data.
This work is based on previous works by N. Kingsbury and I. Selesnick who have obtained several results concerning the dyadic case. Their conclusions are extended to the Mband case. The proposed decompositions (based on two Mband wavelet transforms, the primal and the dual one, operating in parallel) typically introduce a redundancy of a factor 2 (4 in the complex case) and they constitute frames from which we can derive an optimal reconstruction. These new transforms have recently been generalized to biorthogonal and complex cases.
We choose to apply these analysis tools to image denoising, which led us to study the statistical properties of coefficients generated by an Mband dualtree analysis of a widesense stationary random process.
Crosscorrelations between primal and dual wavelets play a major role in our study. Numerical simulations allowed us to validate our theorical results as well as evaluate the influence area of the correlations.
The effectiveness of our decompositions is demonstrated in image denoising. At first, we concentrate on monochannel images to show that Mband dualtree wavelet decompositions bring a significant gain in terms of both objective and subjective quality, with respect to a classical wavelet decomposition as well as a dyadic dualtree wavelet decomposition. Then, we consider multichannel image denoising for which we build a new multivariate estimator based on Stein’s principle. The proposed estimator allows to take into account an arbitrary neighborhood (spatial, intercomponent, interscale...). 

Champs de Markov Bien Nivelés pour la Restauration d'Images ROS avec Préservation du Contraste 
Marc Sigelle Telecom Chief Engineer and Associate Professor ENST Paris 
27/11/2006 14h30 

Résumé (anglais) :
It is now wellknown that when an image containing a bright object within a dark background and overall Gaussian noise corruption is restored using Total Variation (TV) regularization
[Rudin, Osher and Fatemi 1992], a significant loss of grey level contrast between recovered object and background can happen.
We recently showed that TV is the paradigm of those regularization
energy functionals which can be decomposed on level sets, which we called levelable functions [Darbon and Sigelle 2006].
We first show that for gaussian as well as for speckle noise and using levelable priors, the shape of the restored object does not change in a ``statistical'' sense, so that only its brightness decreases whereas that of background increases (loss of contrast).
This extends the results of [Strong and Chan 2003].
We then show how to prevent the result to have this loss of contrast using nicelevelable priors. Some examples are given and appear concluding. This work is described in the research report [Darbon, Tupin and Sigelle 2006].
[Darbon, Tupin and Sigelle 2006]
J. Darbon, M. Sigelle and F. Tupin
A note on NiceLevelable MRFs for SAR image denoising
with contrast preservation.
ENST Research report ENST 2006D006, September 2006.
[Darbon and Sigelle 2006]
J. Darbon and M. Sigelle.
Image Restoration with Discrete Constrained Total Variation Part II:
Levelable Functions, Convex Priors and NonConvex Cases.
Accepted for publication in Journal of Mathematical Imaging and Vision,
March 2006.
[Rudin, Osher and Fatemi 1992]
L. Rudin, S. Osher and E.Fatemi.
Nonlinear total variation based noise removal algorithms.
Physica D. , Volume 60, pp. 259268, 1992.
[Strong and Chan 2003] D. Strong and T. Chan.
Edgepreserving and Scaledependent Properties
of Total Variation Regularization.
Inverse Problems, Volume 19, Number 6 (December 2003), pp. 165187. 

Prediction of forest structure parameters by textural analysis of very high resolution canopy images 
Pierre Couteron, IRD – UMR AMAP (Montpellier) Institut Français de Pondichéry (IFP), India 
17/11/2006 14h30 Coriolis 

Résumé (anglais) :
Predicting stand structure parameters for tropical forests from remotely sensed data has numerous important applications, as estimating aboveground biomass and carbon stocks and providing spatial information for forest mapping and management planning. We shall present the results obtained for both temporarily flooded and terra firme natural forests in French Guiana by applying an approach of "textural ordination" which ordinate or classify canopy image windows of adequate size on the basis of their Fourier spectra.
Whatever the initial information used (IKONOS or digitized photograph contacts) the approach proved able to identify one or two prominent textural gradients, which were interpretable in terms of canopy coarseness/fineness with strong relationships with forest variations as observed from field measurements. Textural indices (i.e. windows scores along the gradients) were good predictors of forest parameters with reasonable estimation errors, and there are good prospects for broadscale implementations. Potential improvements of the approach, which include standardization of spectra originating from images with heterogeneous technical features and fusion of "global" vs. "local" textural measures, will be discussed. 

Waveletbased multichannel image denoising and restoration 
Amel Benazza Associate Professor SUP'COM, Tunis, Tunisia 
13/11/2006 14h30 Coriolis 

Résumé (anglais) :
Multichannel imaging systems provide several observations of the same scene which are often corrupted by additive noise and blurred. In this work, we are interested both in multispectral image denoising and restoration in the wavelet domain.
Concerning the denoising problem, we adopt a multivariate statistical approach in order to exploit the correlations
existing between the different spectral components. Our main contribution is the application of Stein’s principle to build a new estimator for arbitrary multichannel images embedded in Gaussian noise. Simulation tests carried out on optical satellite images show that the proposed method outperforms conventional wavelet shrinkage techniques.
Focusing on the problem of restoration, our approch has two main novelties. At the first step, we show how to combine Mband Wavelet Transforms (WT) with Fourier analysis to restore multicomponent images. At the second step, we point out that the multichannel deconvolution procedure takes advantage of exploiting multivariate regression rules. Simulations experiments carried out on multispectral satellite images indicate the good performance of our method. 

Change Detection in Multitemporal Remote Sensing Images: Basics and Advanced Techniques 
Lorenzo Bruzzone Professor, DIT University of Trento, Italy 
16/10/2006 14h30 Coriolis 

Résumé (anglais) :
This talk presents the basics of change detection in multitemporal remote sensing images and the most recent developments in the context of unsupervised techniques for the automatic analysis of data. After introducing some preliminary concepts related to change detection in images acquired by both multispectral passive sensors and synthetic aperture radars, a novel theoretical framework for change vector analysis (CVA) in multispectral images will be presented and discussed. This framework, which is based on the representation of the CVA in polar coordinates, aims at:
i) introducing a set of formal definitions in the polar domain (which are linked to the properties of the data) for a better general description (and thus understanding) of the information present in spectral change vectors;
ii) analyzing from a theoretical point of view the distributions of changed and unchanged pixels in the polar domain (also according to possible simplifying assumptions);
iii) driving the implementation of proper preprocessing procedures to be applied to multitemporal images on the basis of the results of the theoretical study!
on the distributions; and
iv) defining a solid background for the development of advanced and accurate automatic changedetection algorithms in the polar domain. Examples of application of the proposed framework (and of the related concepts) to real multispectral and multitemporal remote sensing images will be presented. The results obtained confirm the interest of the proposed framework and the validity of the related theoretical analysis. 

The ETATS Project 
Vinciane Lacroix Senior Researcher, Elec. Dept. Royal Military Academy, Brussels, Blegium 
18/09/2006 14h30 Coriolis 

Résumé (anglais) :
In a near future, the National Geographic Institute of Belgium (NGI) will manage its vectorial data ranging from a conceptual scale of 1:10 000 to 1:50 000 in one single database. NGI is setting up a "planning tool" to schedule the data updating process according to the changes that occurred on the field and to compute the uptodate status of the data as information to provide to endusers. The information about the changes will come from various sources, in particular, from remote sensed data. The choice of sensor is such that
(i) the cost of a regular territorial coverage should be affordable;
(ii) the regular territorial coverage should be technically possible;
(iii) its resolution should enable the detection of changes in the builtup area and in the communication network.
As a first step, a visibility test on SPOT5 panchromatic 5m resolution data fused with multispectral data has been performed to assess the possibility for a photointerpretor to detect the buildings and the road network in various type of Belgian landscapes. As this test was positive, an automatic system could be envisaged.
An automatic system to estimate the urbanization changes on the Belgian territory, using SPOT5 data and the National Geographic Institute vectorial database is thus proposed. The images and the vectorial data are first coregistered. Then, the vectorial database is projected and dilated to produce a mask representing the old status of the database. On the other hand, a fusion of two classification processes on the images enables to extract the builtup area and the communication network, generating a mask representing the actual state of the urbanization in the zone. The first process uses simplified Gabor filters to extract structures and texture, while the second is based on the vegetation index and the knowledge about hydrography. The comparison between the two masks provides coarse information on the changes. A validation of the process shows that the system provides about 20% of false alarms for an average of 92% of good detection. 

TwoStep Iterative Shrinkage/Thresholding Algorithms for Total Variation and WaveletBased Image Restoration 
Jose Manuel Bioucas Dias Associate Profesor, IST Technical University of Lisbon, Portugal 
10/07/2006 14h30 Euler Indigo 

Résumé (anglais) :
Image restoration is usually formulated as the minimization of the sum of two convex functions: A quadratic data term and a nonquadratic regularizer (prior in the Bayesian framework). In
recent work, a class of iterative denoising algorithms has been proposed. The denoising operator depends on the regularizer (prior).
Popular regularizers are the 1) Total Variation (isotropic and nonisotropic), 2) the l^p norm, and 3) the pth power of an l^p norm (both 2 and 3 with p greater ou equal to one). The first two classes are usually formulated in the image domain, whereas the former is often formulated in the wavelet domain in applications involving
sparse representations.
The iterative denoising approach is well suited to large scale problems. Its convergence rate is, however, overly slow when the linear operator associated with the data term is illconditioned or illposed. In this talk I will review this class of algorithms and present twostep (also known as second order) versions of the original ones that exhibit a much faster convergence rate. The
underlying motivation behind the twostep versions parallels that of twostep linear methods to solve linear systems of equations.
We show that the proposed twostep iterative scheme converges to a minimum of the underlying optimization problem, for a wide range of regularizers, including those mentioned above. We also give the optimal setting of the parameters that define the algorithm. The
effectiveness of our scheme is illustrated with TV and waveletbased image restoration examples. 

Regionbased extraction and analysis of visual object information 
Ferran Marques Associate Professor UPC Barcelona, Spain 
19/06/2006 14h30 Euler Indigo 

Résumé (anglais) :
Interaction with image database should be as user friendly as possible; that is, it should be as close as possible to the user's normal mode of communication. This concept is the basis for the socalled contentbased image retrieval. To answer user's queries dealing with images containing a specific object, tools to detect the presence of such objects are necessary. We propose in this talk a strategy to detect objects from images. The approach relies on combining two types of models: a perceptual and a structural model. The algorithms that are proposed for both types of models make use of a regionbased description of the image relying on a Binary Partition Tree. Perceptual models link the lowlevel signal description with semantic classes of limited variability. Given that we are dealing here with still images, the temporal evolution of lowlevel descriptors will not be considered to characterize the objects. Structural models represent the common structure of all instances! by decomposing the semantic object into simpler objects and by defining the relations between them. 

Hybrid Imaging 
Yoav Schechner Senior Lecturer, department of Electrical Engineering Technion, Israel 
16/06/2006 14h30 Lagrange Gris 

Résumé (anglais) :
Rather than regarding images as given entities to be processed, richer information is gained by modifying and analyzing the imaging process itself. Such "hybrid imaging" exploits the power to affect both the sensor and the algorithmic components of a vision system. This is especially effective when having multiple sensing inputs.
As examples, we briefly show how mosaicing can be generalized to overcome various radiometric problems (e.g., onuniformity) or obtain high dynamic range, multispectral images. Then, we describe a bioinspired method for recovering visibility in scattering media (haze,water). It is based on independent component analysis (ICA) of multiple frames, aided by a physical model of image formation. This model also easily leads to spatially varying regularization, overcoming inherent distancedependent noise amplification. Finally, we address crossmodal scenarios: we seek correlating features between different inputs (or inputoutput), as done in climatology, medical research, and other fields. Here, we overcome insufficient data by a general mathematical principle, based on a sparsity prior. 

Compressive Sampling 
RobertD. Nowak Associate Professor, Electrical and Computer Engineering University of WisconsinMadison, USA 
29/05/2006 14h30 Coriolis 

Résumé (anglais) :
Compressive sampling (CS), or "Compressed Sensing", has recently generated a tremendous amount of excitement in the image processing community. CS involves taking a relatively small number of nontraditional samples in the form of randomized projections. Such samples are capable of capturing the most salient information in an image. If the image being sampled is compressible in a certain basis (e.g., wavelet), then under noiseless conditions the image can be much more accurately recovered from random projections than from pixel samples. We extended this type of result to show that compressible signals can be accurately recovered from random projections contaminated with noise, in many cases also much more accurately than is possible using an equivalent number of conventional point samples.
I will review the basic theory and methods of CS, and discuss potential applications of CS in imaging and remote sensing. In particular, I will compare CS to conventional imaging by considering a canonical class of piecewise smooth image models. This analysis shows that CS can be advantageous in noisy imaging problems if the underlying image is highly compressible or if the signaltonoise ratio is sufficiently large.


Statistical Analysis of Shapes of 2D Curves, 3D Curves, and Facial Surfaces, by Anuj Srivastava, Florida State University, USA 
Anuj Srivastava Associate Professor, Dept. of Statistics Florida State University, USA 
22/05/2006 14h30 Coriolis 

Résumé (anglais) :
In this talk, I will start by summarizing our previous work on techniques for computing geodesics between planar closed curves,with or without restrictions to arclength parameterizations. Using tangent principal component analysis (TPCA) in this framework, we have imposed probability models on shape spaces and have used them in Bayesian shape estimation and classification. Extending these ideas to 3D problems, I will present a "pathstraightening" approach for computing geodesics between closed curves in R3. The basic idea is to define a space of closed curves, initialize a path between the given two curves, and iteratively straighten this path using gradient of an energy whose critical points are geodesics. Lastly, I will explain how this computation, of geodesics between 3D curves, helps
analyze shapes of facial surfaces. Using level sets of smooth
functions, we represent a facial surface as an indexed collection of facial curves. Then, we compare any two facial surfaces by registering their facial curves, and by comparing shapes of corresponding curves. 

Bayesian Denoising in Oriented and NonOriented MultiScale Pyramids 
Jalal Fadili Associate Professor, Image and Signal Processing ENSI Caen, France 
03/04/2006 14h30 Coriolis 

Résumé (anglais) :
Owing to recent advances in modern harmonic analysis, oriented (e.g. curvelets) and nonoriented (e.g. wavelets) multiscale transforms have proven powerful in sparsely representing a wide class of images and signals (over some smoothness space classes). This work is concerned with bayesian denoising within the context of these sparse representations. A general univariate bayesian prior model, namely the scale mixture of gaussians, on the representation coefficients is introduced and its properties are established. We prove that such a prior is well adapted to capture sparsity of the representations (leptokurticity and heavy tailness). We also shed the light on the relationship between the hyperparameters of this prior and those of the Besov space within which realizations of such a prior are likely to fall. Several bayesian estimators (conditional posterior mean and maximum a posteriori) are also given. All these results are extended to the multivariate case where dependency between neighbouring coefficients in the multiscale pyramid is imposed. Some special cases of the general prior are finally considered (e.g. the bessel K form) and their computational issues are solved. 

On example based segmentation with wavelet features 
Claire Gallagher PhD student, EEE Department Trinity College Dublin, Ireland 
13/03/2006 14h30 Coriolis 

Résumé (anglais) :
The notion of example based image processing in general is one of increasing interest in the signal processing and computer vision community. The idea has long been a keystone in postproduction houses and relies on an initial human generated estimate which is then propagated to other frames. In the image processing community this idea has recently been extended to improve automation. The best examples of this have been in example based texture synthesis (Efros et al. 1999) and Bayesian Matting (Chang et al. 2001,2002). This talk illustrates how the notion of gathering statistics from an image with the help of a human can be used for segmentation. The idea is to integrate the manually labelled example quantitatively into an inference framework for the segmentation of "related but not the same images". Complex wavelets are used to generate features that drive the segmentation process. As an example of the importance of using the complex wavelet, an improved version of the Efr!
os technique is also demonstrated that is much more robust to scale. Example based segmentation holds great potential for putting the user quantitatively into the segmentation loop and the talk concludes with some interesting direction for applications. 

Priorbased Segmentation by Projective Registration and Level Sets, by Tammy RiklinRaviv 
Tammy Raklin Raviv PhD student TAU, Tel Aviv, Israel 
20/02/2006 14h30 Coriolis 

Résumé (anglais) :
Object detection and segmentation can be facilitated by the availability of a reference object.
However, accounting for possible transformations between the different object views, as part of the segmentation process, remains a challenge. Recent works address this problem by using comprehensive training data. Other approaches are applicable only to limited object classes or can only accommodate similarity transformations.
We suggest a novel variational approach to priorbased segmentation, which accounts for planar projective transformation, using a single reference object.
The prior shape is registered concurrently with the segmentation process, without point correspondence. The algorithm detects the object of interest and correctly extracts its boundaries.
The homography between the two object views is accurately recovered as well.
Extending the ChanVese level set framework, we propose a regionbased segmentation functional that includes explicit representation of the projective homography between the prior shape and the shape to segment. The formulation is derived from twoview geometry.
Segmentation of a variety of objects is demonstrated and the recovered transformation is verified. 

Finite states modeling for unsupervised texture segmentation 
Giuseppe Scarpa Ercim PostDoc Academy of Science, Czech Republic 
20/02/2006 15h45 Coriolis 

Résumé (anglais) :
A novel color texture unsupervised segmentation algorithm is presented which processes independently the spectral and spatial information. The algorithm is composed of two parts. The former provides an oversegmentation of the image, such that basic components for each of the texture which are present are extracted. The latter is a region growing algorithm which reduces drastically the number of regions, and provides a regionhierarchical texture clustering. The oversegmentation is achieved by means of a colorbased clustering (CBC) followed by a spatialbased clustering (SBC). The SBC, as well as the subsequent growing algorithm, make use of a characterization of the regions based on shape and context. Experimental results are very promising in case of textures which are quite regular. 

SpaceFusion project: goals and preliminary results 
André Jalobeanu CNRS Researcher LSIIT, ENSPS Strasbourg, France 
13/02/2006 14h30 Coriolis 

Résumé (anglais) :
We propose to develop new multisource data fusion and reconstruction methods. The originality of the project lies in considering data fusion as the estimation of a single model, of arbitrary spatial and spectral resolutions. The model is to be inferred from a number of observations, possibly from different sensors under various conditions.
It is all about reconstructing a geometric and radiometric object that best relates to the observations and integrates all the useful information contained in the initial data.
In astronomical imaging, we will aim at a sharp, correctly sampled, noisefree and possibly superresolved image. In the Virtual Observatory framework for instance, one wishes to combine large numbers of multispectral images from various sources. In planetary imaging or remote sensing, both terrain topography and camera parameters must be taken into account to efficiently combine several images. Therefore, the topography will be included in the model. The object provided by the fusionreconstruction method will be a 3D surface, possibly superresolved regarding both geometry and reflectance.
The first objective is to provide a flexible framework for bandlimited signal reconstruction from multiple data. We obtained promising results in one dimension: a superresolved signal was successfully reconstructed from two blurred and noisy shifted observations. In this framework, the sampling resolution, the geometric distortions, the blur kernel and the regularity of the sampling grid can be arbitrary for each sensor. The method was designed to handle realistic Gauss+Poisson noise and a simple Gaussian Markov chain was used for regularization purposes.
We will also show new insights on the use of this Bayesian approach to infer a highresolution deformation field from two satellite images. 

A Markovian Approach on ForegroundBackgroundShadow Segmentation of Video Images 
Csaba Benedek PhD Student Hungaraian Academy of Science, MTA Sztaki, Hungary 
06/02/2006 14h30 Euler violet 

Résumé (anglais) :
In the talk I introduce a new Markovian model regarding foreground and shadow detection in video sequences. The model works without detailed a priori objectshape information, and is also appropriate for low framerate video sources. I present three novelties: an improved shadow model, a new foreground calculus and a new integration technique of different colortexture features. 

