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 : 2024, 2023, 2022, 2021, 2020, 2019, 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
Image restoration with discrete Level Sets and Total Variation: stochastic aspects	Marc Sigelle ENST, Paris	12/12/2005 14h30 Coriolis
Résumé (anglais) : In this talk we develop a new approach for image restoration with discrete Level Sets and Total Variation (TV) based on a discrete stochastic approach. First we show that Markov Random Field (MRFs) posterior models such as L1 + TV (laplacian noise) and L2 + TV (gaussian noise) can be decomposed and independently optimized on binary level-sets images. This relies on the concept of coupled binary Simulated Annealing. After showing some restoration examples, we discuss the special case of L1 + TV with the "grey-level chessboard" model. Then these results are extended to other types of posterior restoration models. Last we shall prove along the same trend that Perfect Sampling can easily be obtained for a specific subset of these models involving the key notion of convexity. This might in turn pave the way for fast hyperparameter estimation for such a class of models.
Content-based image retrieval using machine learning	Matthieu Cord ENSEA Cergy	05/12/2005 14h30 Coriolis
Résumé (anglais) : The recent domain of image retrieval in large databases has induced a revision of the topics of image processing and pattern recognition. Image retrieval and extraction of visual information from image databases are useful in many applications. In this talk, we focus on content-based retrieval strategies for large image databases. After introducing the global architecture of systems, interactive strategies and relevance feedback are explained. The processing of visual content has emerged as a key area for the application of Machine Learning (ML) techniques. We present different ML techniques that have been applied to the processing of visual information extraction and CBIR applications.
An Approach to Feature Reduction in Hyperspectral Image Classification	Sebastiano Serpico DIBE, Genoa University, Italy	29/11/2005 14h30 Coriolis
Résumé (anglais) : Feature reduction has proved to be an important preprocessing step able to reduce the complexity of hyperspectral image analysis and to allow an improvement in the accuracy of the supervised classification of such images to be obtained. An approach is proposed in this talk aimed at reducing the number of features for the classification of hyperspectral remote sensing images, which is based on an iterative search for local maxima of a distance functional. The proposed search is performed in a discrete space of solutions and is based on progressive moves in the direction of the 'steepest ascent' of the functional. Two kinds of formulation of the feature reduction problem are defined: the former is related to the feature selection problem; the latter is related to a special case of feature transformation, i.e., the case of the averaging of groups of contiguous hyperspectral bands. In both cases, the solution space is a binary (discrete) space, in which a notion of neighbourhood is defined, depending on a size parameter. The convergence properties of the algorithms derived from the proposed approach are analysed. Finally the performances of such algorithms are compared with one each other and with the performances of other well-known feature reduction algorithms by reporting on the experiments with a real hyperspectral image.
Spectral analysis for the generators of stochastic dynamics. Spatial stochastic dynamics	R. Minlos, E. Zhizhina IITP of Russian Academy of Science, Moscow	14/10/2005 14h30 E003
Tracking, mixtures and particles	Patrick Perez Projet VISTA, IRISA Rennes	03/10/2005 Coriolis
Résumé (anglais) : In this presentation, I'll discuss sequential state estimation problems where the filtering distribution is a mixture. Two generic problems, which are of particular interest for visual tracking, fall in this category: (1) sequential estimation of multi-modal filtering distributions, (2) tracking with auxiliary discrete state variables. We shall see that standard filters can be easily extended to handle these problems. In particular, popular sequential Monte Carlo techniques (particle filters) can be readily mobilized. This yields to a clustered particle filter in (1) and to interacting particle filters with no sampling of the auxiliary variable in (2). In both cases, experimental illustration will be provided in the context of colour-based visual tracking.
Segmentation d’images par minimisation de la complexité stochastique avec des modèles paramétrique et non paramétrique du bruit	Philippe Refregier Institut Fresnel, Marseille	19/09/2005 14h30 Coriolis
Résumé (anglais) : We shall present a general image segmentation method adapted to the noise present in the image but which does not need an a priori knowledge of the probability density functions of the grey levels. This method is based on the minimization of the stochastic complexity which leads to a criterion without parameter to be tuned by the user. We shall demonstrate that one can apply this method to different contour descriptors (polygonal active contour, level set implementation and polygonal active grid). We shall analyze on synthetic images the performances of this approach in comparison to those obtained with parametric approaches adapted to the noise model. This technique will finally be illustrated on various real images.
Advances in Statistical Learning from Near-neighbors	Maya Gupta EE Dept, University of Washington, Seattle, USA	09/09/2005 14h30 Coriolis
Résumé (anglais) : We consider the problem of classifying or labeling a test sample based on a database of known samples and their labels. This is a standard statistical learning problem often solved by neural nets, support vector machines, Gaussian mixture models, or decision trees. In this talk research is presented into methods which learn based on near-neighbors; simple examples of this local learning approach are k-nearest neighbor and linear interpolation. Theoretically, we discuss how estimation bias can be reduced by using a convex neighborhood of samples, and by using weights that solve a linear interpolation and maximum entropy objective (LIME). Given weighted neighbors, we show that Bayesian minimum expected risk estimates will significantly outperform maximum likelihood estimates for classification when costs are asymmetric, as is often the case in medical, defense, and non-destructive evaluation applications. Applications include protein structure prediction, the non-destructive evaluation of pipeline integrity, and the automatic creation of custom color enhancements.
Multiscale, Multigranular Statistical Image Segmentation, by Eric D. Kolaczyk, Department of Mathematics and Statistics	Erik Kolaczyk Boston University, USA	11/07/2005 14h30 Coriolis
Résumé (anglais) : In the image segmentation problem, one seeks to determine and label homogeneous subregions in an image scene, based on pixel-wise measurements. Motivated by current challenges in the field of remote sensing land cover characterization, we introduce a framework that allows for adaptive choice of both the spatial resolution of subregions and the categorical granularity of labels. Our framework is based upon a class of models we call "mixlets," a blend of recursive dyadic partitions and finite mixture models. The first component allows for sparse representation of spatial structure at multiple resolutions, while the second enables us to capture the varying degrees of mixing of pure categories that accompany the use of different resolutions. A segmentation is produced in our framework by selecting an optimal mixlet model, through complexity-penalized maximum likelihood, and summarizing the information in that model with respect to a categorical hierarchy. Both theoretical and empirical evaluations of the proposed framework are presented.
Tissue characterization and Detection of Epithelium Dysplasia and Inflammation in-Vitro and in Vivo using Optoelectronically Enhanced Endoscopy Imaging	Fernand Cohen Drexel University, Philadelphia, USA	04/07/2005 14h30 Coriolis
Résumé (anglais) : Malignancies occurring in the epithelium lining the internal surfaces of organs (e.g., aerodigestive tract or colon) typically are discovered late in the course of the disease, usually because they are "hidden" from the physician, and are not usually found on routine physical exam. If there is suspicion of a lesion, endoscopy with biopsy is often necessary. Depending on the location, biopsy may be difficult to obtain, lead to potential complications, or have a low yield. For example, traditional biopsy of the larynx is difficult, because of the sensitivity of the structures and their constant movement. In the esophagus, biopsy of Barrett's mucosa (precancerous tissue), is made more difficult because areas of high malignant potential appear the same as areas of low malignant potential. This makes biopsy in this area fairly low yield. Our work attempts to increase the likelihood of selecting a precancerous sample during biopsy through the use of optoelectronically enhanced endoscopy by developing a tissue characterization technique that is sensitive to changes in the morphology of the tissue as it undergoes changes from normal tissue, to inflammation to oncogenesis. We attempt to establish correlation between changes in parameter values corresponding to changes in the tissue morphology such as cell sizes, shape, density, absorption, index of refraction, etc. Based on the fact that pre-cancerous cells demonstrate shape anisotropy compared to healthy cells, we develop minimally invasive optical techniques that detect the resulting pre-cancerous reflectance signatures. Non-invasive and minimally invasive optical techniques are becoming staples of modern medical technology. An on-chip, high sensitivity probe is created to image the tissue area and placed directly at the tip of an optical fiber, which is em! ployed at the end of an endoscope. In this investigation, we intend to image the tissue area under investigation through the use of scanning with our detector array in order to construct a 3-D "image" of the cellular structure. From this image, we will perform a stochastic decomposition and extract parameter values beyond currently available cell size determination. The parameters will be individually and in combination analyzed to construct a detailed map of pre-cancerous regions. We create a tool that can spatially differentiate between different grades of dysplasia in an inexpensive, portable package. In this talk we concentrate on the theory (Mie scattering), the image decomposition model and report on simulation results, on tissue mimicking phantom, and finally on animal tissue using imaging at one wavelength as well as using white light.
Hierarchical Annealing for the synthesis of porous media images	Simon Alexander University of Waterloo, Canada	20/06/2005 16h00 Coriolis
Résumé (anglais) : While motivated by a particular application, the work described is quite generalizable. Although present in the literature, Simulated Annealing for the synthesis of porous media has met with limited success due to computational costs and practical modelling constraints. An alternative method based on hierarchical annealing will be presented. Inherently multiscale, such approaches may dramatically reduce the computational cost. Energy functions (based on e.g. chord-length distribution) in a hierarchy allow separately treating structures of different length scales, reducing convergence issues for samples with multiple natural scales. Such an approach naturally leads to methods of explicitly multiscale modelling.
Variations on Markovian Quadtree Model for Multiband Image Analysis	Christophe Collet ENSPS-LSIIT, Illkirch	13/06/2005 14h30 Coriolis
Résumé (anglais) : This talk is concerned with the analysis of multispectral observations, provided e.g., by space or ground telescopes. The large amount and the complexity of heterogeneous data to analyse lead to develop new methods for segmentation tasks, which aim to be robust, fast and efficient. Some prior knowledge on the information to be extracted from the original image is available, and Bayesian statistical theory is known to be a convenient tool to take this a priori knowledge into consideration, even if the 'curse of dimensionality' often limits these approaches in a multispectral framework. The main goal of this presentation, consists in showing different processing chains describing the power, the efficiency and the fruitfulness of different hierarchical Markovian modeling based on a quadtree topology. We will see that such modeling allows to deal with a large varieties of data : missing data, multiresolution data, multiband data, strongly noised data. In particular, we show how such approach is general and how this tool is able to face with a large number of various image processing tasks.
Stochastic image modeling, classification, and compression using multitree dictionaries	Ilya Pollack Purdue University, USA	03/06/2005 14h30 E003
Résumé (anglais) : We describe a new framework of multitree dictionaries which includes many previously proposed dictionaries as special cases. We show how to efficiently find the best tree in a multitree dictionary using a recursive tree pruning algorithm. We illustrate our framework through several image compression examples. We then show that it can also be used to construct novel hierarchical stochastic image models called spatial random trees (SRTs). In previously proposed stochastic hidden tree models, the nodes of the tree are labeled with random variables; however, the structure of the tree is fixed and nonrandom. Our key innovation is that both the states and the structure of the tree are random, and are generated by a probabilistic context-free grammar. We describe inference algorithms associated with our SRT models and illustrate them through several image classification and segmentation examples.
Statistical Detection and Estimation of a Biochemical Dispersion in a Realistic Environment	Mathias Ortner University of Illinois, USA	01/06/2005 14h00 E003
Résumé (anglais) : An early detection and estimating the spread of a biochemical contaminant is important in security and pollution monitoring. We present an integrated approach combining the measurements by an array of biochemical sensors with a physical dispersion model and statistical analysis to optimize the system performances. We develop a computational dispersion model of a contaminant in a complex environment through Monte Carlo simulations of reflected stochastic diffusions describing the microscopic transport phenomena due to wind and chemical diffusion. We consider arbitrary geometries and account for wind turbulences. We present examples for two realistic but synthetic scenarios: dispersions in a city environment and an indoor ventilation duct. Using the proposed numerical transport model, we develop an optimal sequential detector. For a fixed false alarm rate, we obtain the detection probability of a substance release as a function of its location and initial concentration. We also compute the expected delay before detection as a function of the source location and initial concentration. Once a contaminant has been detected, localizing the dispersive sources is useful for decontamination purposes and cloud evolution estimation. To solve the associated inverse problem, we propose a Bayesian framework that proved to be powerful for localizing sources with insufficient measurements.
Algorithmes adaptatifs en théorie de l'apprentissage	Albert Cohen Lab. J-L.Lions, Université Pierre & Marie Curie	16/05/2005 14h30 Coriolis
Résumé (français) : Nous considérons ici le modèle de la régression bornée en théorie de l'apprentissage : étant données m réalisations d'une variable (x,y), avec |y| majoré par M, on cherche un estimateur f rendant aussi petit que possible l'erreur |f(x)-y| en moyenne quadratique. L'utilisation des techniques de seuillage par ondelettes est ici rendue difficile par la non-connaissance de la distribution en x, par contraste avec le cadre du débruitage de signaux échantillonnés sur une grille uniforme. Nous présentons et analysons ici une classe d'algorithmes adaptatifs, qui permettent de contourner cette difficulté et qui comme les algorithmes de seuillages sont "universels", au sens où ils ne présupposent pas de régularité particulière sur la fonction de régression que l'on cherche à estimer. Un exemple numérique sera présenté sur des données de terrain.
Segmentation non-supervisée d'images couleurs par un échantillonneur MCMC à sauts réversibles	Zoltan Kato Szeged University, Hungary	25/04/2005 14h30 Coriolis
Résumé (anglais) : Reversible jump Markov chain Monte Carlo (RJMCMC) is a recent method which makes it possible to construct reversible Markov chain samplers that jump between parameter subspaces of different dimensionality. We propose a novel RJMCMC sampler for multivariate Gaussian mixture identification and we apply it to color image segmentation. For this purpose, we consider a first order Markov random field (MRF) model where the singleton energies derive from a multivariate Gaussian distribution and second order potentials favor similar classes in neighboring pixels. The proposed algorithm finds the most likely number of classes, their associated model parameters and generates a segmentation of the image by classifying the pixels into these classes. The estimation is done according to the Maximum A Posteriori (MAP) criterion. The algorithm has been validated on a database of real images with human segmented ground truth.
Event Analysis for the Gamma-ray Large Area Space Telescope	Robin Morris RIACS, NASA Ames Research Center	22/04/2005 14h30 Coriolis
Résumé (anglais) : The Large Area Telescope (LAT) instrument on the Gamma-ray Large Area Space Telescope (GLAST) is currently under construction. It is scheduled for launch in 2007, and its primary mission is an all-sky survey in the energy range 30MeV-100GeV, and its observations should help answer questions about the origins and evolution of the universe. The LAT works by converting gamma-rays in to electron-positron pairs, and then tracking these charged particles as they traverse the instrument, triggering silicon microstrip detectors. The electron and positron also produce secondary emissions, which also trigger the microstrip detectors. I will present work-in-progress on a statistical reconstruction algorithm that incorporates an accurate model of the physics of the detector. It is based on a combination of sequential importance sampling and MCMC, and holds the promise of accurate estimation of the direction and energy of the gamma-ray photons. Results of applying this approach to a simplified representation of the LAT instrument will be presented. Joint work with Dr Johann Cohen-Tanugi, SLAC.
Multiresolution Gaussian Mixtures for Image Analysis	Roland Wilson Warwick University, UK	14/03/2005 14h30 Coriolis
Résumé (anglais) : For years, the image analysis community has used methods which attempt to combine spatial and statistical methods to describe image data of various forms. In recent years, this problem has been compounded by the widespread interest in video data and data from multiple cameras. For the last few years, I have been developing an image representation based on Gaussian mixture modelling, which attempts to address these issues. Cast in a framework of Bayesian estimation, multiresolution GM modelling can be used for data of arbitrary dimensions; the model can be computed efficiently and, crucially, it can be adapted to deal with smooth motions of the underlying co-ordinate system. The talk will give a brief overview of the main theoretical ideas behind MGM and will be illustrated with a number of simple applications, including the representation of motion and stereo imagery.
Imagerie Mathématique : segmentation sous contraintes géométriques - théorie et applications	Carole Le Guyader INSA Rouen	21/02/2005 14h30 E003
Résumé (français) : Dans cet exposé, nous nous intéressons à des problèmes de segmentation d'images sous contraintes géométriques. Cette problématique a émergé suite à l'analyse de plusieurs méthodes classiques de détection de contours qui a été faite. En effet, ces me'thodes classiques (Modèles déformables, contours actifs géodésiques, 'fast marching', etc...) se révèlent caduques quand des données de l'image sont manquantes ou de mauvaise qualité. En imagerie médicale par exemple, des phénomènes d'occlusion peuvent se produire : des organes peuvent se masquer en partie l'un l'autre (ex du foie). Par ailleurs, deux objets qui se jouxtent peuvent posséder des textures intrinsèques homogènes si bien qu'il est difficile d'identifier clairement l'interface entre ces deux objets. La définition classique d'un contour qui est caractérisé comme étant le lieu des points connexes présentant une forte transition de luminosité ne s'applique donc plus. Enfin, dans certains contextes d'étude, comme en géophysique, on peut disposer en plus des doneées d'imagerie, de données géométriques à intégrer au processus de segmentation. Pour pallier ces difficultés, nous proposons ici des modèles de segmentation intégrant des contraintes géométriques et satisfaisant les critères classiques de détection avec en particulier la régularité sur le contour que cela implique.
Chemins minimaux géodésiques sur métrique non scalaire	Hugues Talbot CSIRO, Australia - ESIEE, Marne La Vallée	31/01/2005 14h30 Coriolis
Résumé (français) : Ce séminaire est une extension du travail de Cohen & Kimmel [1] sur les chemins minimaux pour contour actifs, en particulier pour les chemins ouverts, donc utiles pour la segmentation de structures linéiques. On donne l'équation du chemin minimal en métrique tensorielle définie positive, et une solution efficace grâce à une application de l'extension de la méthode de marche rapide (Ordered Upwind Method) due é Sethian et Vladimirsky [2]. On présente une segmentation de réseaux de dendrites neuronales. [1] Global minimum for active contour models: A minimal path approach Laurent D. Cohen and Ron Kimmel. In Proc. IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'96), San Francisco, USA, June 1996. [2] J.A. Sethian and A. Vladimirsky. Ordered upwind methods for static Hamilton-Jacobi equations. Proc. Natl. Acad. Sci. USA 98/20: 11069-11074 (2001).