La reconstruction dense de surfaces 3D a partir d'images 2D est un probleme particulierement mal pose. Notre equipe a choisi d'utiliser une approche bayesienne. Cela permet d'introduire des contraintes sur l'objet a reconstruire, afin de stabiliser la solution, au moyen d'un modele statistique de la surface.

Nous presentons d'abord un modele possible de surface, qui peut etre utilise pour decrire des asteroides, aussi bien que pour modeliser des surfaces planetaires. Le support consiste en un reseau triangulaire de topologie spherique, appele surface subdivisee, car il est obtenu par subdivision recursive d'un polyedre initial. Sur ce support nous definissons un modele statistique, qui decrit aussi bien la geometrie de l'objet que sa reflectance. En introduisant une transformee en ondelettes sur le reseau subdivise, nous construisons ainsi un modele multiechelle qui permet de prendre en compte le comportement fractal des surfaces naturelles.

Ensuite, nous presentons une nouvelle methode de rendu recemment developpee, qui permet de generer des images a partir d'une surface 3D. Elle est nettement plus precise que les algorithmes de rendu existants, tout en tenant compte des occlusions et des ombres. Elle genere egalement, pour chaque pixel, les derivees de l'intensite par rapport a tous les parametres. Cette technique, couplee au modele multiechelle, permet en principe de reconstruire la surface inconnue a partir de plusieurs observations prises par differentes cameras. Nous presenterons notre contribution par rapport aux travaux deja effectues dans l'equipe en matiere de reconstruction de terrain 3D.

We present a new image segmentation algorithm based on a tree-structured binary MRF model. The image is recursively segmented in smaller and smaller regions until a stopping condition, local to each region, is met. Each elementary binary segmentation is obtained as the solution of a MAP estimation problem, with the region prior modeled as a MRF. Since only binary fields are used, and thanks to the tree structure, the algorithm is quite fast, and allows one to address the cluster validation problem in a seamless way. In addition, all field parameters are estimated locally, allowing for some spatial adaptivity. To improve segmentation accuracy, a split-and-merge procedure is also developed and a spatially adaptive MRF model is used. Moreover, a recent implementation, that allows to define different and independent binary fields on unconnected region of the same class, is presented. This last solution presents improved capacity of detail description, while no additional computing time is required.

The seminar will consist of two parts. I will start with presenting a method for extracting small tracks from the remotely sensed imagery. The method is Bayesian and the MAP estimate is sought by the Gibbs sampler coupled with simulated annealing. Second half of the seminar will be devoted to the results of evaluation of the Gibbs sampler ability to improve the quality of the initial classification (by QDA or ICM) of the multispectral noisy images.

Discrete Wavelet transforms are usually realised using the filterbank framework. The polyphase matrix of such filterbanks can be factored into lifting steps, to achieve the lifting framework of wavelet transforms. The lifting framework provides an useful and flexible tool for constructing new wavelets from the existing ones. In this talk, we present the designing of spatially adaptive wavelet transforms using the lifting framework. The high pass and low pass filters of the resulting in wavelet transform are chosen spatially adaptively by considering the statistics of the underlying signal. The perfect reconstruction can be achieved without coding any side information regarding to the filter selection. The performance of such transforms in lossless, lossy and scalable image and video coding is presented.

This talk concerns two problems of continuous-parameter annealing: large-scale pixellated models, and small-state parameterized models (such as Markov point processes).

The estimation of large-scale images from sparse and/or noisy data is highly-developed, however although estimates are optimum under some criterion, they do not represent a typical or representative sample of the system being studied, which may be desired for purposes of visualization, further analysis, Monte Carlo studies etc. Instead, what is required is that we find a random sample from the posterior distribution, a much more subtle and difficult problem than estimation, and, crucially, one which cannot be formulated as an optimization problem. I will present a little-known property of multiscale statistical models to formulate a posterior sampler, exact in the case of Gauss-Markov random fields, and approximate for other distributions.

For parameterized models, widely-scattered problems such as formant tracking, boundary estimation and phase-unwrapping can all be approached as the annealed minimization of continuous parameters. In virtually all such annealing problems the Metropolis sampler is used, where the effectiveness of the annealing is highly-dependent on a user-formulated query function. I will propose an alternative - to use the Gibbs sampler, which requires no query, but which requires the difficult sampling of nonparametric, multimodal distributions.

Image segmentation is the art of describing image data, which tend to be highly complex, in terms of simpler entities, such as regions of homogeneous gray level, colour or texture. This amounts to attaching an integer label to each pixel in an image, representing its class. In the last decade or so, it has become widely accepted that the problem can be formalised as a maximisation of a posteriori probability, based on a stochastic image model, such as a Markov Random Field (MRF). While this puts segmentation on a firm footing, it raises a significant issue in terms of computation: how can one possibly maximise the posterior probability over the huge number of possible image segmentations, given a set of data? In recent years, two methods have found widespread use: stochastic simulation samples from the posterior distribution of the image model; multiresolution methods exploit the self-similarity of image data to solve a sequence of successively finer approximations to the problem. It has occurred to a number of authors that these two approaches might be usefully combined in a multiresolution MRF. The work we have done using these models has thrown up interesting results in both the theory and practice of image segmentation. In the talk, I will examine the segmentation problem and present some of our results.

One of the major problems in geographical information systems (GISs) consists in defining strategies and procedures for a regular updating of land-cover maps stored in the system databases. This crucial task can be carried out by using remote-sensing images regularly acquired by space-born sensors in the specific investigated areas. Such images can be analysed with automatic classification techniques in order to derive updated land-cover maps. However, at the operating level, such techniques are usually based on supervised classification algorithms. Consequently, they require the availability of ground truth information for the training of the classifiers. Unfortunately, in many real cases, it is not possible to rely on training data for all the images necessary to ensure an updating of land-cover maps that is as frequent as required by applications. In this seminar, advanced classification techniques for a regular updating of land-cover maps are proposed that are based on the use of multitemporal remote sensing images. Such techniques are developed within the framework of partially supervised Bayesian approaches and are able to address the updating problem under the realistic but critical constraint that, for the image to be classified (i.e., the most recent of the considered multitemporal dataset) no ground truth information is available. Two different approaches are considered. The first approach is based on an independent analysis of the information contained in each single image of the considered multitemporal series; the second approach exploits the temporal correlation between pairs of images acquired at different times in the classification process. In the context of these approaches, both parametric and non-parametric classifiers are considered. In addition, in order to design a reliable and accurate classification system, multiple classifier architectures composed of partially supervised algorithms are investigated. Experimental results obtained on a real multitemporal and multisource dataset are presented that confirm the effectiveness of the proposed system.

Since the discovery of the cosmic microwave background (CMB) radiation in 1965 by Penzias and Wilson, a number of missions have been planned to measure the CMB including the ESA satellite PLANCK. CMB radiation is of interest from a number of aspects: 1) it is the most important evidence for the hot big-bang model, 2) it provides us a picture of the universe in its very early moments 3) the anisotropies in it provide the seed map of the universe of today 4) It provides us information about the fundamental constants of the universe. Unfortunately, the task of measuring CMB is not an easy one, since radiation measurements of the sky contain contributions from various sources from our galaxy such as syncrotron, galactic dust and free-free emission as well as extragalactic radio sources. In this talk, I will present our efforts at ISTI-CNR to separate various radiation sources in astronomy images. We adopt a source separation rather than a noise elimination approach since we value the information in other sources as well. I will start with our work using independent component analysis (ICA) as a fully blind technique and then will move into more informed techniques such as independent factor analysis (IFA) which assume a generic source model and a full Bayesian approach using MCMC. Talk will end with the description of the future activities we plan.

Le tatouage des images consiste à insérer de manière secrète et indélébile de l'information d'indexation ou de protection contre les copies au sein du signal. Dans un certain nombre de situations, le signal à protéger subit des transformations géométriques importantes (par exemple, une version de l'image imprimée sur du papier, ou une image projetée sur un écran dans une salle de cinéma). Nous établirons un état de l'art sur les méthodes permettant de retrouver l'information de tatouage malgré ces déformations. Nous exposerons des nouvelles techniques originales basées sur le lien entre l'information de tatouage et des caractéristiques essentielles du signal, invariantes lors des transformations géométriques.

The light microscope is unique in its ability to image cells and display morphology and molecular localization at sub-cellular resolutions. As such, it served biological research for centuries, based on human understanding of microscopic scenes. Digital imaging added the quantitative capability, mainly based on fluorescent immunostaining, which enabled to study molecular localization and displacements inflicted by various experimental manipulations and drugs. Today modified cDNA allows to express in cells inherently fluorescent tagged proteins and follow them as cells respond to stimuli. With cDNA libraries, the design of cell-based large scale experiments open ways to identify the protein networks that underlie complex cellular mechanisms. But such experiments depend on automated acquisition and analysis of microscope images from many samples. Various analysis methods applied to cell-based assays will be described. Specialized aspects of biological quantitative imaging will be discussed, and examples of segmentation, quantification and comparison of multicolor and time-lapse images will be shown.

Dans ce séminaire nous allons étudier la convenance sur l'utilisation de la matrice de Pascal dans le dessin des filtres numériques. Cette matrice nous permet d'effectuer la transformation de la fonction de transfert H(s) pour obtenir sa version discrète H(z) d'une manière simple. Egalement la matrice inverse de Pascal est utilisée pour transformer H(z) dans H(s) sans avoir besoin de calculer le déterminant du système.

The presentation is dedicated to emerging aspects of stochastic image modeling in critically sampled and overcomplete transform domains for image restoration and denoising and consists of three main parts. In the first introduction part, we briefly present stochastic image processing (SIP) group and its current research activity. The second part of presentation reviews robust image restoration for radar and radiometry imaging systems.

First, we consider the general problem formulation and corresponding solution based on a penalized maximum likelihood estimate. The relationship to robust M-estimators and maximum a posteriori probability methods will be indicated for different stochastic image priors and noise distributions. We also consider non-coherent imaging systems based on sparse antenna arrays and demonstrate some practical results. The third part of presentation is dedicated to the important problem of stochastic image modeling in the transform domains. We consider the! state-of-the-art stochastic image models and different classes of multiresolution transforms. We will focus on important class of non-stationary image models in different applications and will analyze its main shortcomings based on an example of estimation-quantization (EQ) model. Finally, we will introduce a novel edge process (EP) model for critically sampled and overcomplete domains and demonstrate its main advantages. In conclusion, some upper bounds for the EP model performance in image denoising application will be demonstrated.

Sensor networks have emerged as a fundamentally new tool for monitoring spatial phenomena. This talk will describe a theory and methodology for estimating inhomogeneous fields using wireless sensor networks. Inhomogeneous fields are composed of two or more homogeneous regions (e.g., constant-valued, smoothly-varying, stationary Gaussian, etc.) separated by boundaries. The boundaries, which correspond to abrupt spatial changes in the field, are non-parametric 1-d curves or 2-d surfaces (in a 2-d or 3-d field, respectively). The sensors make noisy measurements of the field, and the goal is to obtain an accurate estimate of the field at some desired destination (typically remote from the sensor network). The presence of boundaries makes this problem especially challenging. There are two key questions: 1. Given n sensors, how accurately can the field be estimated? 2. How much energy will be consumed by the communications required to obtain an accurate estimate at the destination? Theoretical upper and lower bounds on the estimation error and energy consumption will be discussed. A practical strategy for estimation and communication will be presented. The strategy, based on a hierarchical data-handling and communication architecture, provides a near-optimal balance of accuracy and energy consumption.

The nonparametric multiscale algorithms presented here are powerful new tools for photon-limited signal and image denoising and Poisson inverse problems. Unlike traditional wavelet-based multiscale methods, these algorithms are both well suited to processing Poisson data and capable of preserving image edges. The recursive partitioning scheme underlying these methods is based on multiscale likelihood factorizations of the Poisson data model. These partitions allow the construction of multiscale signal decompositions based on polynomials in one dimension and multiscale image decompositions based on platelets in two dimensions. We originally developed platelets for medical image reconstruction problems, and more recently we have successfully applied them to problems in astronomical imaging. Platelets are localized functions at various positions, scales and orientations that can produce highly accurate, piecewise linear approximations to images consisting of smooth regions separated by smooth boundaries. Polynomial- and platelet-based maximum penalized likelihood methods for signal and image analysis are both tractable and computationally efficient. Simulations establish the practical effectiveness of these methods in applications such as Gamma Ray Burst intensity estimation and medical and astronomical image reconstruction; statistical risk analysis establishes the theoretical near-optimality of these methods.

Ce seminaire presente une chaine de traitement pour l'extraction des batiments a partir d'images aeriennes. Dans un premier temps, nous nous focalisons sur la d´etection des batiments rectangulaires qui sont le type de construction le plus repandu. Nous etendons notre methode aux batiments plus complexes, qui peuvent etre decomposes en plusieurs rectangles. Les rectangles obtenus permettent d'ameliorer la reconstruction 3D du Modele Numerique d'Elevation (MNE). La segmentation du MNE et de l'ortho-image permet l'extraction du sur-sol. Nous calculons un critere de ressemblance entre chaque region et leur meilleur rectangle associe. Pour les batiments complexes, nous proposons un algorithme de division des regions. Le decoupage optimise iterativement notre critere de ressemblance. L'approche est illustree sur des donnees synthetiques et reelles. L'estimation des structures rectangulaires n'est pas correctement localisee ni dimensionnee. Nous presentons un modele parametrique deformable qui permet d'ameliorer ces caracteristiques. Les estimations rectangulaires finales sont utilisees avec leur altitude extraite du MNE pour obtenir une scene 3D precise.