Thanks to recent progress in the iterative decoding of concatenated codes, several new fields of investigation have appeared. In this paper, we present a first approach of the iterative decoding of Reed-Solomon (RS) product codes:

Cet article contient les développements de l'exposé présenté en séance plénière lors du 11ème Symposium IFAC/IFORS sur Identification and System Parameter Estimation (SYSID'97) qui s'est tenu à Kitakyushu, Japon, du 8 au 11 Juillet 1997. On décrit les principes fondamentaux et des exemples d'application réels d'une théorie unifiée de conception de traitements embarqués pour la détection précoce et le diagnostic de pannes de composants. De tels traitements sont nécessaires pour la surveillance de systèmes en vue de leur maintenance conditionnelle. On met en évidence le fait que la principale tâche pour la conception de tels algorithmes consiste en la sélection d'une fonction d'estimation convenable. Les méthodes d'identification du maximum de vraisemblance, des moindres carrés, des variables instrumentales, et de type sous-espaces, sont discutées dans cette perspective.

On considère le problème de la détection et du diagnostic de pannes modélisées par des changements de la structure propre d'un système dynamique linéaire. Dans cet article sont décrits et analysés de nouveaux algorithmes de détection et diagnostic de telles pannes, basés sur les méthodes récentes d'identification par sous-espaces et sur l'approche statistique locale. L'application à la surveillance vibratoire des structures mécaniques et machines tournantes est discutée. Ces méthodes sont comparées conceptuellement à un autre algorithme basé sur la méthode d'identification par variables instrumentales, et proposé précédemment par deux des auteurs.

La conduite et la maintenance de procedes industriels necessitent 'analyse de leur comportement tel qu'il est percu par les divers capteurs. s'agit essentiellement de detecter, localiser et diagnostiquer toute par rapport a un comportement de reference. Nous presentons des statistiques du traitement du signal et des methodes de l'intelligence permettant la generation des alarmes et leur interpretation. les illustrons par des exemples tires d'applications industrielles. nous efforcons de montrer la complementarite de ces approches.

The aim of this paper is to validate a compression scheme applied on a medical data base of digitized wrist radiographs. The compression scheme adapts to local statistical properties of the images. The diagnostic quality the reconstructed images is evaluated using a ROC protocol involving five experts. The results of this evaluation enable us to validate the scheme on this data base with a compression ratio of 40 (0.2 bits per pixel).

This article presents an overview of several measures for speaker recognition. These measures to second-order statistical tests, and can be expressed under a common formalism. formulations of these measures are given and their mathematical properties are studied. their basic form, these measures are asymmetric, but they can be symmetrised in various way. measures are tested in the framework of text-independent closed-set speaker identification. appears to be a factor of improvement, especially when little speech material is. The use of some of the proposed measures as a reference benchmark to evaluate the complexity of a given database under a given protocol is finally suggested as a conclusion this work.

In the field of mobile robotics, range sensing is the most important point of any autonomous system. The navigation of a mobile robot depends on the accuracy of its perception system. The localization of an indoor autonomous robots with maps produced by 2D laser rangefinders is a real problem. Indoor environments are mainly made of connected or non-connected planes (e.g. walls, doors, desks...). Data consist in a set of points belonging to straight lines. Thus polygonal approximation is the more suitable model for those data. Therefore, classical methods are not convenient for this kind of data. This paper proposes a new method of segmentation, including polygonal approximation and fusion of segments. We also present experimental results that demonstrate successful map building.

The paper deals with a fuzzy classification algorithm. It is applied to the interpretation of aerial images and to simulated data, in a supervised and partially supervised environment. membership grade assigned to the classes allows a correct classification even if the classes overlap.

As an image cannot be well segmented by only one segmentation method, paper deals with a new paradigm, called segmentation tree. structure allows the application of several operators on the same image. principle is the cooperation between high level and low level processes in image analysis. segmentation process is controlled by the interpretation quality.

L'article presente un systeme de vision qui, par une comparaison locale entre plusieurs segmentations/interpretations determine la meilleure segmentation d'une image. deux points importants sont la construction d'un arbre qui contient les segmentations possibles d'une image et la comparaison des resultats de segmentation, a partir de l'interpretation des regions. donnees symboliques (dediees a une application,ici les images aeriennes) et numeriques sont exploitees conjointement. capacite d'apprentissage des objets d'une scene par des techniques de reconnaissance des formes, ainsi que la capacite de generaliser l'apprentissage a des objets non appris sont etudiees.

Description d'un modele deformable surfacique applique a l'extraction surfaces du coeur (endocarde et epicarde)

We present classification methods, which have been adapted to the pattern recognition problem in the domain of aerial imagery. Three methods based on different mecanisms have been developped : rule-based system, neural network and fuzzy classification. Our contributions are the adaptation of these methods to the concrete case of aerial images, and the quantitative comparison between the three types of methods.

Cette communication présente un exemple d'application du logiciel SynDEX d'aide à l'implantation d'algorithmes sur des architectures multiprocesseurs. L'application choisie est une méthode de compression d'images, elles est portée sur un réseau de transputers.

This work is a contribution to adequation between algorithm and architecture. It presents an example of application made with SynDEX, a software environment to implement signal processing or automatic algorithms on multiprocessors networks. It is shown that a conditional data flow graph is enough to realize the implementation.

This paper presents a new scheme for fractal image compression based on adaptive triangulation. Such a partition is computed on an initial set of points with a split and merge algorithm in a grey level dependent way. The is thus fully flexible and returns a limited number of blocks good compression ratios. Moreover, a second original approach is the of a classification step based on a modified version of the Lloyd (vector quantization) in order to reduce the encoding complexity. -the vector quantization algorithm is implemented on pixel histograms directly from the triangulation. The aime is to reduce the number of comparisons the two sets of blocks involved in fractal image compression by keeping the best representative triangles in the domain blocks set. Quality results are achieved at rates between 0.25-0.5 bpp depending on the of the original image and on the number of triangles retained.

Cet article donne des généralités sur la fusion de données dans le cadre d'un problème de reconnaissance. De façon générale, sont exprimés les besoins propres à ce problème. Nous définissons notamment quelles sont les informations à considérer et les traitements que celles-ci doivent subir pour mener à bien le problème traité. Nous montrons également l'intérêt d'utiliser une solution multi-sources, en particluier, dans certains cadres d'application où les informations disponibles sont souvent imparfaites. Dans ce contexte, un panorama des méthodes ou théories, utilisées comme formalisme de modélisation, est proposé en montrant les inconvénients et les avantages de chacune d'elles.

Cet article s'intéresse à la fusion de sources définies sur des référentiels non exhaustifs différents. C'est le cas par exemple des sources S1 et S2 définies respectivement sur les référentiels H1,H2,H3 et H2,H3,H4. Une source prise isolément est insatisfaisante car ne permet de discerner toutes les hypothèses. La fusion de ces sources apparaît alors comme une solution intéressante pour obtenir un discernement d'un plus grand nombre d'hypothèses et d'autre part de tirer profit de la complémentarité de discernement sur les hypothèses communes. Toutefois, les processus de combinaison fournis par les différents formalismes ne sont pas applicables dans un tel cas de fugure. On propose donc différentes solutions permettant de réaliser une telle fusion. Comme cadre d'application privilégié de ce problème, nous avons choisi la reconnaissance dans le cas d'un apprentissage incomplet.

Cet article s'intéresse à la fusion de sources définies sur des référentiels non exhaustifs différents. C'est le cas par exemple des sources S1 et S2 définies respectivement sur les référentiels H1,H2,H3 et H2,H3,H4. Dans ce cas, une source prise isolément est insatisfaisante car elle ne considère pas toutes les hypothèses. La fusion de ces sources apparaît alors comme une solution intéressante pour obtenir un discernement d'un plus grand nombre d'hypothèses et de tirer profit de la complémentarité de discernement sur les hypothèses communes. Toutefois, les processus de combinaison fournis par les formalismes existants ne sont pas applicables ici. On propose donc différentes solutions pour réaliser une telle fusion. Un cadre privilégié de ce problème est la reconnaissance dans le cas d'un apprentissage incomplet.

In this paper, we study the advantage of the use of the mathematical morphology tools in grey levels in the detection of microcalcifications in digital mammograms. will detail the novel technique developed to separate the microcalcifications which appear as clusters in the picture.

La lithotritie consiste à fragmenter les calculs (lithiases) urinaires et biliaires afin de permettre leur évacuation par les voies naturelles. Le succès du traitement est lié entre autres à la capacité à faire coïncider la lithiase et le foyer du générateur d'ondes de choc acoustiques. L'objectif du travail présenté est de suivre la lithiase dans les images fournies par l'échographe intégré au lithotriteur, afin de déplacer automatiquement le générateur d'ondes suivant les mouvements de la lithiase. L'algorithme choisi s'appuie sur une mesure de corrélation et utilise une stratégie de recherche multigrille à resserrage de maille.

Dans ce travail nous abordons les problemes de l'estimation de mouvement et du recalage de signaux et d'images sous hypothese de conservation des niveaux d'intensite. Dans une premiere partie nous introduisons une nouvelle hypothese de rigidite du mouvement et une pyramide multiéchelle adaptée permettant d'ecrire une nouvelle équation de contrainte du mouvement valide aux échelles grossieres. Nous en deduisons des conditions suffisantes de convergence d'un algorithme itératif d'estimation/ recalage au sens de la disparition du flot optique résiduel et du recalage de l'image d'étude vers l'image cible, en dimension quelconque. Dans une seconde partie nous discutons la mise en oeuvre de l'algorithme et tentons l'approche par ensembles de niveaux en dimension 1 et 2. Ensuite nous obtenons des résultats satisfaisants en utilisant l'algorithme du gradient conjugue. Nous illustrons l'approche à l'aide de nombreux exemples synthétiques ou réels en dimension 1 (y compris sur paire stéréo) et 2.

The objective of the project is to design a system of interpretation of images issued from different sources. needs the dynamic generation of optimal sequences of image processing algorithms, adapted to the current context. original approach for a pertinent qualification of existing image processing algorithms, based on Dempster-shafer theory is proposed.

In this paper, we focus on the TM image processing dealing with the left ventricle's motion in the short axis view. Our goal is an automatic extraction of the endocardial contour over time, in order to compute its motion parameters. We propose a novel algorithm combining spatial discontinuity detection and temporal smoothing. The discontinuities are extracted from each TM-line , by means of a robust detector which measures the degree of the local non-stationarity of the signal. The ventricle's cavity is approximately extracted by an adaptive thresholding so as to help the selection of the endocardial contours from the detected set of discontinuity lines. These contours are then approximated by cubic splines. Instantaneous velocities of the heart walls are finally computed by derivating the obtained temporally smooth curves.

The paper presents some experiments for speaker verification using a single layer perceptron. The speaker information is transmitted in a speech segment by the dependent and independent phoneme information. The experiments confirm the possibility of speaker verification using the speaker information contained in the short time spectrum.

Part 1: THE LMS THEORY OF ADAPTIVE TRANSVERSAL FILTERS; Introduction - The Optimal Filter; Adaptivity and Recursivity: the LMS Filter; The LMS Case with Independent Input Vectors; A More Realistic Model to Prove Convergence; Transient Error and the Speed Convergence; Fluctuations in Steady-state and Misadjustment. 2: IMPLEMENTATION CONSIDERATIONS; Complexity and Precision of the LMS algorithm; The Simplified Sign Algorithms. 3: TRACKING THE TIME-VARIATIONS; Adaptive Filtering in Presence of Time Variations; Tracking Bounded Time Variations; Tracking Random Time Variations; The Digital Phase Tracking Loop. 4: SOME LMS ADAPTIVE RECURSIVE (HR) FILTERS: Recursive Filtering and Prediction in Transmission; LMS Recursive Predictors; Stability of Adaptive Recursive Prediction.

This paper provides a general theory for the problem of tracking the variations of an filter with an adaptive filter: LMS, normalized LMS, signed LMS, RLS etc... the unrealistic assumptions of independence, infinitesimal forgetting rate, or noise input, we evaluate the tracking performance thanks to a general Markov which splits the adaptive filter deviation into its transient, noise and lag . In the specific random walk case, LMS and RLS exhibit a trade-off forgetting rate compromizes between noise and lag misadjustments. Their tracking achievements be in favour of LMS or RLS. With an order 1 Markov model, it is shown for LMS that trade-off between lag and misadjustment occurs only if the output noise in the bandwidth this Markov process is smaller than the output signal power -up to a fixed factor. it is preferable to switch off adaptation after the acquisition period.

The problem under consideration is to separate an unknown mixture of unknown signals. The key property that makes separation possible is statistical of the sources, which means both second order and higher order properties. This paper is concerned with self-adaptive such system. with a recursive network controled with the Hérault-Jutten updating rule, present two other linear structures, the direct one and the mixed direct/ one, which do not raise realizibility problem. For two sources it is that the three structures always possess at least one stable stationary that is a separating state, under a condition of globally sub-gaussian . The performance comparison between them shows that the mixed structure the best one because, for the same separation achievement in steady-state, its speed has the best robustness against the variations of the mixture.

In this paper we propose a new adaptive algorithm to separate a linear mixture of using an extended Anti-Hebbian rule. This solution can be viewed as a gradient way to minimize certain output high order statistics. The is modular: it is decomposed into parallel and independent sub-systems. one is capable of extracting one source with negative kurtosis out of the, provided the number of observations is greater or equal to the number sources and provided it is appropriately initialized.

In the classical methods of unsupervised source separation, the a priori hypothesis independence of sources. In certain applications, there is some additional on the sources (statistics, distributions, alphabet...). It is the case discrete sources with known alphabet. Then we can improve separation. of adaptation is done according to some known algorithm, e.g. thanks an extended anti-Hebbian algorithm, provided there are not less sensors than. As soon as the separation performance index has reached some preassigned, a second part which involves the output decision error is introduced in the. In a noiseless environment, this method allows complete cancellation of state adaptation fluctuations and perfect source recovery.

L'utilisation de structures linéaires et de lois d'adaptation mettant en jeu des linéarités d'ordre quatre a été à l'origine de nombreux résultats intéressants séparation non supervisée de sources. Ces performances n'auraient pas été sans que des hypothèses a priori sur les distributions des sources et indépendance soient faites. Dans cet article, on utilise une structure à base neurones non linéaires bouclés. En rajoutant l'hypothèse de sources binaires, laé de séparation se voit améliorée et la robustesse du système vis-à-vis des des paramètres du mélange et de la présence d'un bruit additif est augmentée.

In applications digital recursive filter coefficients have no distinct values, therefore the test of an entire family of polynomials is required in order to be sure of the filter stability. The edge theorem by Bartlett, Hollot and Lin states that a polytope is stable, if and only if, the exposed edges are stable. In this paper, this last condition is transformed into an equivalent one that can be tested in a finite number of arithmetic operations and from which an algorithm is derived. It is shown that the condition, which has been established, is optimal i.e., it can neither be avoided, nor simplified.

On s'intéresse au diagnostic de pannes dans des systèmes hybrides numériques/symboliques. L'objectif est la conception systématique d'algorithmes de diagnostic utilisant conjointement les graphes de défaillances provenant d'études de fiabilité et le traitement des données des capteurs embarqués. On propose une nouvelle approche, basée sur les idées des et l'algorithme de Viterbi. On en donne des résultats expérimentaux pour un système mécanique simulé. On discute des extensions possibles pour tenir compte d'un contrôleur discret d'une part et de la concurrence entre pannes d'autre part.

This paper derives an identification solution of the optimal linear predictor of ARMA type, as a time varying lattice of arbitrarily fixed dimension, for a whose output signal only is known. The projection technique used leads an hereditary algorithm which is the adaptive extension to raw data of our results on lattice realization from given autocorrelation functions. produces a minimum phase linear model of the signal whose n-th order

A new nonlinear technique is developped by means of infinite impulse response functionals. It yields the projection onto the closed class of finite series with separable kernels of arbitrary degree k. If the original model is bilinear, the computation of the kernel is finitely recursive higher moments of degree 2k.

This paper is concerned with the problem of separating independent non-Gaussian . This is done adaptively maximizing a contrast function based on fourth- cumulants of the (mixed) observations. The first class of solutions involves first stage where the signal vector is adaptively whitened. In order to implement the second stage the proper separating task, new contrast functions are proposed, when all the source kurtosises have the same sign. These contrasts only self-cumulants of the outputs. The second class of solutions requires single separating stage. However, the associated contrasts involve cross-cumulants addition to self-cumulants. They essentially apply to correlated vectors with powers (rather than to white vectors). The resulting adaptive one-stage two-stage systems achieve satisfactory separation performance independently of statistics of sources and of the kind of linear mixture.

This paper deals with an automatic registration method of multidate and multisensor remote sensing images as preprocessing in a multichannel consensual segmentation scheme. This two step method consists in a geometrical correction followed by a residual translation compensation performed by detecting the crosscorrelation maximum between multidate imgaes. This second step is preceded by an edge detection for SPOT and prefiltered ERS multisensor images.

Le sujet de cet article est la revue des techniques visant à estimer le champ de vitesses apparentes en évitant de lisser les discontinuités de mouvement. La première partie de l'article présente les méthodes s'appuyant sur l'hypothèse selon laquelle les discontinuités de mouvement coïncident spatialement avec certaines frontières photométriques. La deuxième partie rapporte les méthodes de segmentation du champ estimé courant en régions homogènes au sens du mouvement. La troisième partie s'intéresse aux méthodes qui réalisent simultanément l'estimation du champ de vitesses et la détection des ruptures locales de continuité de ce champ. Ceci permet d'inhiber localement le lissage, dans les zones susceptibles de contenir une discontinuité de mouvement. inhibition peut être obtenue par introduction d'un

When displaying three-dimensional (3D) magnetic-resonance angiography (MRA) images by maximum-intensity-projection (MIP), some vessels may be invisible, partially or completely, although they are visible in individual slices. Indeed, the MIP technique tends to increase the mean level of the background, since it

The aim of the presented work is to improve the visibility of the vessels in magnetic resonance angiography (MRA) images. The proposed technique is designed to filter out noise, to enhance small low-intensity vessels and to preserve larger structures (e.g. aneurisms) with their exact boundaries. It is based on anisotropic smoothing, applied to the data volume before the maximum intensity projection (MIP). The first stage of the filter extracts the local orientation of the vessel. The second stage performs a non-linear smoothing in the vesselís local direction so as to avoid blurring its boundaries. The noise level reduction, the contrast enhancement and the resulting small vessel visibility improvement achieved by this filter are illustrated on dynamic contrast-enhanced subtraction (DCS) MRA images of the lower limbs. These results show that the orientation-based filtering has great benefits for the visibility of the low-intensity vessels in MIP images of MRA data.

The method tries to take the best of two segmentation approaches : a high resolution edge detector and a texture measure to decide whether to suppress the portion of edge or to keep it and prolong it. The method is composed of two steps, both of them exploiting distances between textures. The first step erases the false contours in texture areas. The second step closes the remaining contours.

This paper presents a low complexity block turbo decoder for product codes. This new decoder, which has been derived from the near-optimum block turbo decoder described in [1], is a very good compromise between complexity and performance. For performance evaluation, we have considered the [BCH(64,57,4)]2 product code transmitted over a Gaussian channel using a QPSK modulation. The complexity of the new block turbo decoder is about ten times less than that of the near-optimum block turbo decoder for a coding gain degradation of only 0.7 dB.