Search

National audience; Cet article présente une mesure de voisement basée sur le calcul du signal analytique. Cette mesure peut être utile pour plusieurs applications concernant le traitement de la parole. Par exemple, considérant la reconnaissance automatique de la parole, elle pourrait être incorporée dans les vecteurs acoustiques communément utilisés, comme par exemple les Mel Frequency Cepstral Coefficients (MFCC) et leurs deux premières dérivées, ceci pour améliorer les performances du système. La base de données TIMIT est segmentée manuellement en phonèmes : c'est pourquoi l'évaluation de la mesure développée est effectuée sur cette base. L'information de voisement est déduite de cette segmentation. Il est montré dans cet article que la segmentation automatique voisé/non-voisé obtenue en utilisant la méthode décrite dans les sections suivantes et la segmentation manuelle voisé/non-voisé fournie dans TIMIT sont très similaires.

Les travaux présentés dans ce manuscrit ont pour but le développement de méthodes de prise de décisions optimales (statiques ou séquentielles) ou de traitement de signaux par des méthodes d'appren- tissage automatique. Ils répondent cependant à un certain nombre de contraintes qui ont été imposées par l'objectif de prendre en compte la présence de l'humain dans la boucle de décision ou de traitement ou même comme générateur des signaux que l'on analyse. La présence de l'humain rend la nature des données que l'on traite assez imprévisible et, à une même situation en apparence, des décisions diffé- rentes peuvent être prises. Dans les domaines de l'apprentissage statistique, du contrôle optimal ou du traitement du signal ceci se traduit par la nécessité de gérer l'incertain, de traiter le caractère stochastique du système ainsi que la non-stationnarité de celui-ci. Ainsi, les décisions que l'on considère optimales peuvent dépendre de manière aléatoire de la situation mais de plus, cette dépendance peut varier avec le temps. Des méthodes d'optimisation convergeant vers une solution globale ne sont donc pas adaptées mais des méthodes permettant d'apprendre au fil de l'eau et de poursuivre l'évolution de la solution optimale seront préférées. Par ailleurs, dans le cas où les décisions résultent en une action sur le monde extérieur, il est né- cessaire de quantifier le risque pris en accomplissant ces actions, particulièrement si ces actions doivent avoir un impact sur l'humain. Ceci passe par une estimation de l'incertitude sur le résultat des actions possibles et la sélection en conséquence de ces actions. Une autre implication est qu'il ne sera pas tou- jours envisageable de tester toutes les actions possibles pour en estimer les effets puisque ces actions peuvent ne pas être acceptables pour l'humain. Ainsi, il faudra apprendre à partir d'exemples de situa- tions imposées ce qui se traduit par une phase d'inférence utilisant les informations observables pour en déduire des conséquences sur des situations que l'on ne peut observer. Les travaux exposés dans ce manuscrit apportent des contributions théoriques permettant de tenir compte de ces contraintes et des applications à des problèmes concrets imposant ces contraintes seront exposées.

International audience; Brain Computer Interfaces (BCI) are systems enabling hu- mans to communicate with machines through signals generated by the brain. Several kinds of signals can be envisioned as well as means to mea- sure them. In this paper we are particularly interested in even-related brain potentials (ERP) and especially visually-evoked potential signals (P300) measured with surface electroencephalograms (EEG). When the human is stimulated with visual inputs, the P300 signals arise about 300 ms after the visual stimulus has been received. Yet, the EEG signal is often very noisy which makes the P300 detection hard. It is customary to use an average of several trials to enhance the P300 signal and reduce the random noise but this results in a lower bit rate of the interface. In this contribution, we propose a novel approach to P300 detection using Kalman ltering and SVMs. Experiments show that this method is a promising step toward single-trial detection of P300.

En Imagerie par Résonance Magnétique (IRM) à rehaussement de contraste, la segmentation des structures internes du rein est nécessaire pour une étude de la fonction rénale par compartiment. Pour éviter une segmentation manuelle fastidieuse, deux méthodes (semi-)automatiques, utilisant un algorithme de quantification vectorielle visant à regrouper les pixels rénaux d'après leurs vecteurs temps-intensité, sont proposées et validées sur des données réelles.

International audience; Black Blood Fast Spin Echo imaging of the heart is usually performed during mid-diastolic rest. This is a direct consequence of the long inversion time required to suppress the blood signal, which is constrained by the T(1) of the blood, and of the heart rate. To overcome these constraints, and to acquire black blood images in the end-systolic rest period, a new approach is introduced aiming at adaptively predicting the best time to prepare and acquire MR signals. It is based on a RR interval prediction algorithm and on a cardiac cycle model. The proposed method was applied to 14 healthy volunteers and is compared to a simple alternative method using a fixed delay and to the standard black blood imaging method for imaging in the mid-diastolic rest period. Results show that the proposed method offers an increased robustness in terms of trigger delay error and image quality compared to the tested simple alternative. Also, it has been shown by qualitative analysis done by an experienced observer that the right ventricle, especially the thin right ventricle free wall, is better depicted with our method than with the standard mid-diastolic rest acquisition.

International audience; User simulation has become an important trend of research in the field of spoken dialog systems because collecting and annotating real man-machine interactions with users is often expensive and time consuming. Yet, such data are generally required for designing and assessing efficient dialog systems. The general problem of user simulation is thus to produce as many as necessary natural, various and consistent interactions from as few data as possible. In this paper, is proposed a user simulation method based on Bayesian Networks (BN) that is able to produce consistent interactions in terms of user goal and dialog history but also to simulate the grounding process that often appears in human-human interactions. The BN is trained on a database of 1234 human-machine dialogs in the TownInfo domain (a tourist information application). Experiments with a stateof- the-art dialog system (REALL-DUDE/DIPPER/OAA) have been realized and promising results are presented.

International audience; In recent years reinforcement-learning-based approaches have been widely used for management policy optimization in spoken dialogue systems (SDS). A dialogue management policy is a mapping from dialogue states to system actions, i.e. given the state of the dialogue the dialogue policy determines the next action to be performed by the dialogue manager. So-far policy optimization primarily focused on mapping the dialogue state to simple system actions (such as confirm or ask one piece of information) and the possibility of using complex system actions (such as confirm or ask several slots at the same time) has not been well investigated. In this paper we explore the possibilities of using complex (or hybrid) system actions for dialogue management and then discuss the impact of user experience and channel noise on complex action selection. Our experimental results obtained using simulated users reveal that user and noise adaptive hybrid action selection can perform better than dialogue policies which can only perform simple actions.

Edited by Mohamed Najim

International audience; Data collection and annotation are generally required to design or assess spoken dialogue systems. Yet, this is a very time consuming and expensive process. For these reasons, user simulation has become an important trend of research in the field of spoken dialogue systems. The general problem of user simulation is thus to produce as many as necessary natural, various and consistent interactions from as few data as possible. In this paper, we propose a user simulation method based on Bayesian networks (BN) that is able to produce consistent interactions in terms of user goal and dialogue history. The model as been introduced in previous work but parameters were hand-tuned and it was assessed in the framework of automatic learning of optimal dialogue strategies. In this paper, the BN is trained on a database of 1234 human-machine dialogues in the TownInfo domain (a tourist information application). Experiments with a state-of-the-art dialogue system (REALL-DUDE/DIPPER/OAA) have been realized and results in terms of dialog statistics are presented.

International audience; In clinical applications, cardiac gated sequences are commonly used for heart imaging. In practice, we simply wait a fixed time called Trigger Delay (TD) after the R-wave has been detected to acquire data in the chosen cardiac phase. The main problem of this method is that the TD is constant and does not take into account the physiological variability such as instantaneous heart rate changes during breath-hold or free breathing. Another issue is that we cannot use a TD shorter than the preparation time. Consequently when a cardiac gated sequence with a long preparation time is used, only diastolic images are achievable. Double Inversion Recovery Fast Spin Echo (DIR-FSE), resulting in black-blood images of the heart [1], is such a sequence since the inversion time (TI) needed to cancel the blood signal is around 500ms. Systolic view of the heart could also be of clinical interest, especially during end-systolic phase when the cardiac volume is minimum and constant (average TD = 300ms, average duration 60ms [2]). The primary aim of this work was to assess a robust method to acquire black blood FSE in end-systolic phase. For this purpose, the DIR preparation has to be launched before the R-wave in the previous cardiac cycle. A RR interval prediction is then needed and heart rate variability has to be accounted in order to position the acquisition window properly. A new general adaptive method is described here that overcomes the above listed limitations in prospectively cardiac gated sequences with long preparation time using a predictive model. This method has been applied on end-systolic black blood FSE images.

Au carrefour des mathématiques et des sciences pour l'ingénieur, le traitement statistique du signal trouve aujourd'hui des applications dans tous les secteurs d'activité. Ce livre, qui suppose le lecteur maîtrisant déjà des notions d'algèbre linéaire et de calcul des probabilités, lui offre la possibilité d'acquérir les principales notions utilisées dans la littérature scientifique plus spécialisée. Il traite des principaux problèmes peu nombreux rencontrés dans un grand nombre d'applications, en insistant sur les hypothèses et les démonstrations mathématiques permettant de justifier qu'un outil ou une méthode statistique est optimal. L'ouvrage donne un cadre mathémtique très précis et rigiureux dans lequel l'ingénieur en quête d'innovation peut se repérer pour choisir une direction de recherche afin d'améliorer un système quelconque. Conçu pour des élèves en troisième année de grande école d'ingénieurs et des étudiants en deuxième année de master des facultés des sciences, le livre peut également être lu avec intérêt par des ingénieurs ou des chercheurs en activité.

L'imagerie par résonance magnétique (IRM) du coeur reste difficile. Les solutions retenues dans la pratique clinique pour limiter les problèmes liés aux mouvements sont l'apnée et la synchronisation cardiaque. En imagerie cardiaque, la séquence Double Inversion Recovery Fast Spin Echo (DIR-FSE) est couramment utilisée [1]. Cependant, cette séquence, synchronisée avec l'activité cardiaque, nécessite un temps de préparation long (TI). Le temps entre l'onde R et l'acquisition (TD) est donc également long. Par conséquent, cette séquence n'est utilisée qu'en diastole (figure 1-a). De plus, les variations de rythme cardiaque instantanées (temps entre 2 ondes R consécutives) ne sont pas prises en compte. Afin de passer outre ces limitations, nous proposons deux techniques pour acquérir les séquences DIR-FSE en fin de systole (relaxation iso-volumétrique, volume cardiaque minimum et constant). Ces acquisitions en fin de systole devraient nous permettre de mieux voir le ventricule droit en raison de sa contraction. Les deux méthodes proposées seront comparées entre elles en terme de robustesse. Elles seront également comparées à la méthode conventionnelle en terme de visualisation du ventricule droit.

International audience; The kernel trick is a well known approach allowing to implicitly cast a linear method into a nonlinear one by replacing any dot product by a kernel function. However few vector quantization algorithms have been kernelized. Indeed, they usually imply to compute linear transformations (e.g. moving prototypes), what is not easily kernelizable. This paper introduces the Kernel-based Vector Quantization (KVQ) method which allows working in an approximation of the feature space, and thus kernelizing any Vector Quantization (VQ) algorithm.

International audience; Due to heart motion, cardiac MRI is made difficult and image acquisitions have to be synchronized with heart activity to suppress cardiac motion artifacts. Electrocardiogram (ECG) is the most accurate tool for this purpose, and Triggering consists in synchronizing MR sequences on the R-waves. The complex MR environment worsens ECG acquisition conditions because of the static magnetic field (Hall Effect), Radio Frequency (RF) and fast switching magnetic field gradients. Many hardware developments have been achieved so as to limit these undesirable effects [1], but gradient artifacts are still a problem and signal processing is required. Two ways of research have been followed; (a) first way consists in building a MR specific QRS detector [2], which is based on the vectocardiogram (VCG) the 3D representation of heart activity. This method does unfortunately not provide a clear ECG and is unable to process patient with low VCG amplitude. (b) Secondly a real-time gradient artifact suppression method has been designed. This method is based on adaptive signal processing [3] and achieves real-time accurate denoising which enables triggering. This method requires gradient signals information, which is a major drawback as the connection to the MR system is rarely available. In this paper a new real-time gradient artifact correction, which does not require any connection to the MR system, is presented. The magnetic field pulse signals will be provided by a new specifically MR designed 3D Hall probe, integrated on a low cost 0.35\μm CMOS technology [4, 5].

The invention relates to a method of determining motion-related features (Fm) pertaining to the motion of an object (1), which method comprises obtaining a sequence of images (f1, f2,..., fn) of the object (1), processing at least parts of the images (f1, f2,..., fn) to extract a number of first-order features from the images (f1, f2,..., fn), computing a number of statistical values pertaining to the first-order features, combining the statistical values into a number of histograms (HLC, HG), and determining the motion- related features (Fm) for the object (1) based on the histograms (HLC, HG). Furthermore, the invention relates to a method for performing motion classification for the motion of an object (1) captured in a sequence of images (f1, f2,..., fn).; Such a motion classification method comprises determining motion-related features (Fm) for the images (f1, f2,..., fn) using the described method of determining motion-related features (Fm), and using the motion-related features (Fm) to classify the motion of the object (1). The invention also relates to a system (3) for determining motion-related features (Fm) pertaining to an object (1) in an image (f1), and to a system (4) for performing motion classification for the motion of an object ( 1 ).

A method and a motion recognition system is disclosed for recognizing a motion pattern of at least one object by means of determining relative motion blur variations around the at least on object in an image or a sequence of images. Motion blur parameters are extracted from the motion blur in the images, and based thereon the motion blur variations are determined by means of determining variations between the motion blur parameters.

International audience; In dynamic contrast-enhanced magnetic resonance imaging (DCE- MRI), segmentation of internal kidney structures like cortex, medulla and pelvo-caliceal cavities is necessary for functional assessment. Manual segmentation by a radiologist is fairly delicate because images are blurred and highly noisy. Moreover the different compartments cannot be delineated on a single image because they are not visible during the same perfusion phase for physiological reasons. Nevertheless the differences between temporal evolution of contrast in each anatomical region can be used to perform functional segmentation. We propose to test a semi-automated split and merge method based on time-intensity curves of renal pixels. Its first step requires a variant of the classical Growing Neural Gas algorithm. In the absence of ground truth for results assessment, a manual anatomical segmentation by a radiologist is considered as a reference. Some discrepancy criteria are computed between this segmentation and the functional one. As a comparison, the same criteria are evaluated between the reference and another manual segmentation.

Pour traiter et stocker le volume croissant de données issues de capteurs d'images multispectrales et hyperspectrales embarqués dans les satellites de nouvelle génération, le développement d'algorithmes de compression capables d'exploiter les redondances spectrale et spatiale de ces données est nécessaire. Une précédente étude a montré que, pour des schémas de compression associant des décompositions en ondelettes 2D (pour réduire la redondance spatiale) à des transformations linéaires adaptables (pour réduire la redondance spectrale), la transformation spectrale optimale, à hauts débits, peut être calculée par des algorithmes d'Analyse en Composantes Indépendantes (ACI) modifiés. Dans le cadre de nos travaux de recherche, nous avons poursuivi cette étude en nous focalisant sur deux points essentiels. Le premier a consisté à rechercher une matrice à base d'ACI à coefficients fixes qui ne s'adapte plus à l'image, dans le but de réduire la complexité de codeurs utilisant des transformations optimales. Pour y parvenir, nous avons calculé une transformation quasioptimale exogène sur une base d'apprentissage constituée d'images issues d'un même capteur, puis testé cette transformation sur une base de test différente de la base d'apprentissage. Les performances en compression des transformations exogènes à base d'ACI ont été évaluées et comparées à celles obtenues par la KLT en utilisant comme codeur le VM9 (Verification Model version 9), qui est compatible avec la Partie 2 du standard JPEG2000. Les premiers résultats obtenus ont été complétés par une autre étude portant sur des images Hyperion (voir l'exposé de I. P. Akam Bita et al.) Le second point a consisté à rechercher un codeur par construction d'arbres de zéros 3D bien adapté aux transformations à base d'ACI. Nous avons commencé par développer un codeur n'utilisant que des arbres de zéros 2D, de type SPIHT, qui gère des images 3D avec une allocation quasioptimale entre composantes. Ce SPIHT modifié, que nous appellerons SPIHT2D dans la suite a été conçu pour gérer des images 3D avec le meilleur compromis que nous avons obtenu entre complexité de l'algorithme et performances. Disposant de ce codeur SPIHT2D de référence, nous avons cherché des structures par arbres de zéros 3D et comparé les performances de codeurs utilisant ces structures (ezw3d, spiht_3D, spihtN_3D et spihtN_3D2, la valeur de N est un paramètre) associés ou non à un codeur arithmétique avec modélisation de contextes.

In previous works, we defined the Optimal Transform Code (OTC) assuming high rate coding and using the asymptotical Bennett approximation of the rate. We showed that the OTC gives the optimal linear transform of a multicomponent image compression scheme which consists in applying a linear transform that adapts to the encoded image for reducing the spectral redundancy and a fixed 2-D Discrete Wavelet Transform (DWT) per component for reducing the spatial redundancy. The performances in terms of rate vs PSNR (Peak of Signal to Noise Ratio) are very attractive when evaluated with the Verification Model version 9 of the JPEG2000 committee which is a JPEG2000 codec (coding-decoding). The transform in OTC performs better than the Karhunen Loeve Transform (KLT). The drawback of the OTC is its high computing complexity, since the optimal linear transform must be computed for each encoded image. In order to implement the OTC in an on- board satellite real-time codec system, we propose to pass round the problem of computing complexity by learning only one fixed transform with the OTC algorithms from a set of images instead of computing a new transform for each image. We call the fixed transform computed in this way an exogenous quasi-optimal linear transform. In this paper, we focus the study on hyperspectral images. Our set of images is constituted of ten Hyperion3 hyperspectral images. We have separated the VNIR and the SWIR bands (since they are obtained with two different sensors on- board) and we just focus on the VNIR spectral bands.

International audience; During the last decade, research in the field of Spoken Dialogue Systems (SDS) has experienced increasing growth. However, the design and optimization of SDS is not only about combining speech and language processing systems such as Automatic Speech Recognition (ASR), parsers, Natural Language Generation (NLG), and Text-to-Speech (TTS) synthesis systems. It also requires the development of dialogue strategies taking at least into account the performances of these subsystems (and others), the nature of the task (e.g. form filling, tutoring, robot control, or database search/browsing), and the user's behaviour (e.g. cooperativeness, expertise). Due to the great variability of these factors, reuse of previous hand-crafted designs is also made very difficult. For these reasons, statistical machine learning (ML) methods applied to automatic SDS optimization have been a leading research area for the last few years. In this paper, we provide a short review of the field and of recent advances.

Dans cet article, un cadre théorique pour la simulation et l'optimisation automatique de systèmes de dialogues vocaux entre homme et machine par le biais d'un apprentissage non-supervisé de stratégies est proposé. Ce cadre s'appuie sur une description probabiliste de la communication parlée entre homme et machine. Il permet de s'inscrire dans le cadre des processus décisionnels de Markov et de faire usage de l'apprentissage par renforcement pour rechercher une stratégie optimale de manière indépendante de la tâche. Deux applications concrètes du cadre proposé aux cas du remplissage de formulaire et de l'interrogation de bases de données sont données afin d'en démontrer les utilisations possibles.

National audience

Edited by Mohamed Najim

National audience

International audience; Double Inversion Recovery Fast Spin Echo (DIR-FSE) sequences, resulting in black-blood images, are widespread in cardiac exams [1] as high resolution and great contrast between heart cavities and walls are achieved. For these sequences, DIR pulses are generally played on the R wave and the inversion time (TI) needed to cancel blood signal (~500ms) only allows diastolic view of the heart. Cardiac contraction is complete at end systole so myocardium can be better observed during this short phase. To perform systolic acquisitions, DIR pulses have to be placed before the R wave in the previous heart cycle and RR interval prediction is mandatory. Moreover, as systole is generally shorter than diastole, shorter echo trains are desirable, resulting in longer acquisition time incompatible with breath hold. A method which combines (i) heart rate prediction [2], (ii) respiratory motion estimation [3] and (iii )motion compensated construction [4] is presented. It allows free breathing black-blood systolic imaging and has been tested on five subjects.

International audience; Cardiac MRI is still challenging due to the perpetual heart motion. In order to avoid cardiac motion artifacts, acquisitions are synchronized with heart activity, generally by triggering on R-waves of the Electrocardiogram (ECG). Double Inversion Recovery Fast Spin Echo (DIR-FSE) sequences, resulting in black-blood images, require furthermore specific timing. For such acquisitions, inversion time (TI) is needed to cancel blood signals (~500ms) [1], making acquisitions in the first 500ms of the cardiac cycle impossible. Black-blood imaging allows then only diastolic view of the heart. Moreover heart rate (HR) variability has been shown to be important during breath-holds [2] making cardiac synchronization even more complicated. In this paper, a RR interval prediction method has been implemented, which permits to launch DIR pulses before R waves and thus enables black blood systolic imaging.

International audience; In this study, we examine whether it is reliable to use all successive shocks from one patient for the development of an outcome predictor model to discriminate "Success" versus "No success". Vector of predictors !v are extracted from time and non-linear dynamics domains and a Gaussian Mixture Model-based bayesian classifier, with probability density estimated by the Expectation-Maximization algorithm, is applied in order to detect shocks with "Success" according to the probability P(!v =Success). A cross-validation analysis is performed independently on 136 first shocks (Group1) and 382 second and later shocks (Group2). At 5 s post-shock, an Organized Rhythm (OR) is considered as "Success" and Ventricular Fibrillation (VF) is defined as "No success". A decrease in performance of discrimination of OR versus VF between Group1 and Group2 is observed with an Area Under the ROC Curve of 0.82 and 0.65, respectively. This corroborates the current hypothesis that within-patient correlation affects defibrillation outcome prediction accuracy.

National audience; L'imagerie par résonance magnétique (IRM) du coeur reste difficile. Les solutions retenues dans la pratique clinique pour limiter les problèmes liées aux mouvements sont l'apnée et la synchronisation cardiaque. En imagerie cardiaque, la séquence Double Inversion Récupération Fast Spin Echo (DIR- FSE) est couramment utilisée [1]. Cependant, cette séquence, synchronisée avec l'activité cardiaque, nécessite un temps de préparation long (TI) ce qui contraint le temps entre l'onde R et l'acquisition (TD) à être long. Par conséquent, cette séquence n'est utilisée qu'en diastole. De plus, les variations de rythme cardiaque instantané (temps entre 2 ondes R consécutives) ne sont pas prises en compte (Figure 1-A et 1-B). Afin de passer outre ces limitations, nous désirons mettre en oeuvre une technique adaptative pour utiliser les séquences DIR-FSE en systole (volume cardiaque minimum et constant). Ceci a deux buts : 1. étudier et diagnostiquer certaines maladies en systole et 2. étudier les propriétés magnétiques du coeur en systole et diastole.