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
Hermite projection method in image processing and analysis	Andrey KRYLOV Professor, Head of the Laboratory of Mathematical Methods of Image Processing Lomonosov Moscow State University	05/12/2011 14h30 salle Coriolis
Résumé (anglais) : The method is based on the expansion of the data into series of Hermite functions being the eigenfunctions of the Fourier transform. The set of these real orthogonal functions is full in L2; the functions have a good localization both in "time" and "frequency". It can be considered as a "localized" substitution of the trigonometric basis. The functions are closely related with "Gaussian derivatives". Short mathematical justification will be given. Fast Hermite projection method will be described. The method will be illustrated with examples in image filtering, image foveation, iris data parameterization, image keypoints extraction.
Déconvolution non-supervisée et myope	Jean-François GIOVANNELLI Professeur Equipe Signal et Image, LAPS/IMS, Université de Bordeaux 1	21/11/2011 14h30 salle Coriolis
Résumé (français) : La première partie de la présentation concerne la construction d'une classe de champs aléatoires composites à potentiel L2-L1 (type Huber, non quadratique mais convexe) dont la caractéristique principale est de posséder une fonction de partition explicite et simple. On explicite ainsi la dépendance des lois de probabilités vis-à-vis des hyperparamètres et on rend possible l'estimation de ces paramètres (inversion non-supervisée) dans les problèmes inverses et notamment en déconvolution d'images. On inclut également l'estimation conjointe des paramètres instrument (inversion myope). L'inférence bayésienne et les techniques d'échantillonnage (MCMC de la forme Gibbs et Metropolis-Hastings) fournissent alors une estimée à la fois de l'objet, des hyperparamètres et des paramètres instruments.
New trends in optical imaging for cells and tissues investigations	Hervé RIGNEAULT Director of Research Institut Fresnel, Marseille	17/10/2011 14h30 salle Coriolis
Résumé (anglais) : We will review some new methodology and technology developments in optical microscopy. Special interest will be given to wide field phase imaging and to molecular imaging as viewed by stimulated Raman techniques.
Segmentation conjointe d’images	Stéphane DERRODE Associate Professor Institut Fresnel, Marseille	26/09/2011 14h30 salle Coriolis
Résumé (français) : Un problème important en traitement du signal consiste à restaurer un processus inobservable à partir d’un processus observé. Lorsque l’on souhaite analyser conjointement deux séries de données, ie y1 et y2, on peut soit construire un modèle de mélange vectoriel aboutissant à une classification unique x pour les deux séries d’observations, soit construire une modèle « couplé » ou « conjoint » qui fournit deux classifications x1 et x2 distinctes mais statistiquement liées. L’objet de l’exposé est de présenter en détail un modèle conjoint correspondant au modèle bien connu de mélange probabiliste. Pour ce modèle, nous présenterons deux critères bayésiens de segmentation et un algorithme EM pour l’estimation des paramètres, rendant ainsi la méthode de classification conjointe entièrement non supervisée. Des résultats d’expériences concernant des données et des images simulées seront présentés, permettant d’évaluer la méthode comparativement à d’autres. Enfin, plusieurs applications sur des images multi-spectrales et multi-temporelles seront présentées, qui intègrent des lois d’attache aux données paramétriques non-gaussiennes. Nous terminerons l’exposé en montrant comment cette idée peut s’adapter aux modèles de Markov cachés.
Continuous Convex Relaxation Methods for Image Processing: Optimal Solutions and Fast Algorithms	Xavier BRESSON Assistant Professor City University, Hong Kong	11/07/2011 14h30 salle Euler bleu
Résumé (anglais) : This talk will introduce recent methods to compute optimal solutions to fundamental problems in image processing. Several meaningful problems in image processing are usually defined as non-convex energy minimization problems, which are sensitive to initial condition and slow to minimize. The ultimate objective of our work is to overcome the bottleneck problem of non-convexity. In other words, our goal is to “convexify” the original problems to produce more robust and faster algorithms for real-world applications. Our approach consists in finding a convex relaxation of the original non-convex optimization problems and thresholding the relaxed solution to reach the solution of the original problem. We will show that this approach is able to convexify important and difficult image processing problems such as image segmentation based on the level set method and image registration. Our algorithms are not only guaranteed to find a global solution to the original problem, they are also at least as fast as graph-cuts combinatorial techniques while being more accurate.
Compressed Sensing with Poisson Noise	Rebecca WILLETT Assistant Professor Duke University	01/07/2011 14h30 salle Euler bleu
Résumé (anglais) : Compressed sensing has profound implications for the design of new imaging and network systems, particularly when physical and economic limitations require that these systems be as small and inexpensive as possible. However, several aspects of compressed sensing theory are inapplicable to real-world systems in which noise is signal-dependent and unbounded. In this work we discuss some of the key theoretical challenges associated with the application of compressed sensing to practical hardware systems and develop performance bounds for compressed sensing in the presence of Poisson noise. We develop two novel sensing paradigms, based on either pseudo-random dense sensing matrices or expander graphs, which satisfy physical feasibility constraints. In these settings, as the overall intensity of the underlying signal increases, an upper bound on the reconstruction error decays at an appropriate rate (depending on the compressibility of the signal), but for a fixed signal intensity, the error bound actually grows with the number of measurements or sensors. This surprising fact is both proved theoretically and justified based on physical intuition.
Approches non locales et régularisation optimisée par coupure minimale pour le débruitage de données radar	Florence TUPIN Professeur Département TSI Telecom ParisTech	20/06/2011 14h30 salle Euler bleu
Résumé (français) : Dans cette présentation, nous présentons deux familles d'approches pour régulariser les images radar. Nous présentons tout d'abord les approches non-locales et leur reformulation dans un cadre probabiliste. Celle-ci réécrit le problème comme un problème d'estimation au sens du maximum de vraisemblance pondéré et permet de traiter les données pour lesquelles un modèle de distribution du bruit est disponible. Dans le cas de l'imagerie radar, cela permet d'aborder aussi bien des données en amplitude qu'interférométriques ou polarimétriques. Nous présentons ensuite les approches par minimisation d'énergie combinant un terme de vraisemblance et un terme d'a priori sur la régularité de la solution cherchée. Ces énergies peuvent être optimisées de façon exacte ou approchée par la recherche de la coupe minimale dans un graphe approprié. Nous montrerons qu'elles peuvent apporter un cadre de travail pour les données interférométriques et notamment la combinaison de données multi-canal.
Evaluation and design of linear reconstruction methods with the frequency error kernel	Laurent CONDAT Senior CNRS Researcher GREYC, Caen	09/05/2011 14h30 salle Euler bleu
Résumé (anglais) : Linear reconstruction in a shift-invariant space of a function or its derivatives from uniform linear measurements is a classical low-level operation in signal and image processing. Applications include all resampling tasks and gradient evaluation for volume rendering. In this general framework, we show that the frequency error kernel is a particularly useful tool to evaluate and design reconstruction schemes. In the last part of the talk, I will present linear resampling methods between lattices having the same density, which are fully reversible while being fast and of high quality.
Spectral-spatial classification of hyperspectral images using hierarchical optimization	Yuliya TARABALKA Postdoctoral Researcher NASA Goddard Space Flight Center, USA	21/04/2011 14h30 salle Coriolis
Résumé (anglais) : Hyperspectral imaging provides rich spectral information for every pixel in a particular scene, hence increasing the ability to distinguish physical structures in the scene. However, a large number of spectral channels presents challenges for image classification. While pixelwise classification techniques process each pixel independently without considering information about spatial structures, further improvements can be achieved by the incorporation of spatial information in a classifier, especially in areas where structural information is important to distinguish between classes. In this talk, we will present novel strategies for spectral-spatial classification of hyperspectral images. In particular, we will discuss classification techniques using hierarchical optimization. We will discuss new dissimilarity measures between image regions and convergence criteria. We will show that the new techniques improve classification accuracies and provide classification maps with more homogeneous regions, when compared to previously proposed methods.
Some Structure Methods for the Analysis of Images	Gui-Song XIA Research assistant TSI Dept. Telecom ParisTech	11/04/2011 14h30 salle Coriolis
Résumé (anglais) : This presentation focuses on the use of geometric methods in the analysis of images and textures. Investigations in the work either rely on the level lines of images or on the somehow dual and less structured notion of gradient orientation. The first part of this presentation presents a shape-based texture analysis scheme grounded on the topographic map, a tree organization of the level lines of images. Also using the topographic map representation, the second part of this presentation develops a general approach for the abstraction of images, the aim of which is to automatically generate abstract images from realistic photographs. The subject of the last part of this presentation is the detection of junctions in natural images. The approach relies on the local directions of level lines through the orientation of image gradient.
Automatic Detection and 3D Reconstruction of Buildings and Roads from Aerial and Satellite Images	Beril SIRMACEK Research Fellow German Space Agency (DLR)	28/03/2011 14h30 salle Coriolis
Résumé (anglais) : High resolution optical aerial and satellite images provide valuable information to researchers. With their availability, there has been much interest to extract man-made objects from such imageries. Among these, detection of structures such as buildings and road segments play crucial roles especially for municipalities, government agencies, rescue teams, military and other civil agencies. Manual extraction of this information is too much tedious and prone to errors. Therefore, automated methods are needed to extract information from these images. Unfortunately, the solution is not straightforward by using well-known image processing and pattern recognition methods. Hence, advanced algorithms should be developed to detect buildings and roads automatically. Besides, detecting building and roads, their three-dimensional reconstruction also has very high importance, since it can help to detailed three-dimensional urban region monitoring, change detection analysis, and disaster simulations. In this talk, we introduce our novel approaches for automatic detection of buildings and roads from panchromatic satellite images using local invariant features. These local invariant features help us to develop robust mathematical techniques which can work under various imaging conditions (like different illumination conditions, different scale, looking angle, etc.). Besides, we also introduce methods for detailed three-dimensional urban region reconstruction. Our experimental results on a diverse data set including high resolution panchromatic Ikonos and WorldView-1 satellite images indicate practical usefulness of our approaches.