72 research outputs found
UNE APPROCHE VERS LA COMPREHENSION AUTOMATIQUE DES TEXTES ARABES DESTINEE POUR LES SYSTEMES DE QUESTION-REPONSE
No full textIn order to obtain the DOCTORAT Informatique Par M me. WITED BAKARI AN APPROACH TO THE AUTOMATIC UNDERSTANDING OF ARAB TEXTS INTENDED FOR QUESTION-ANSWER SYSTEMS Supported on March 15, 2018, before the jury composed of: Mr. Rafik BOUAZIZ Professor of Higher Education at FSEG-Sfax President Mr. Mohamed JEMNI Professor of Higher Education at ENSI-Tunis Rapporteur Mrs. Nadia ESSOUSSI Professor of Higher Education at ISG-Tunis Rapporteur Mr Faiez GARGOURI Professor of Higher Education at ISIM-Sfax Examiner Mr Mahmoud NEJI Lecturer at FSEG-Sfax Thesis DirectorLe travail de cette thèse sřinscrit dans un cadre général du traitement automatique de lalangue naturelle (TALN). Il se situe dans le contexte de la recherche dřinformations précises.Nous nous intéressons plus précisément aux systèmes de question-réponse. En effet, cettethèse a pour but de proposer une nouvelle approche pour la question-réponse arabe. Cetteapproche sera implémentée en un système de question-réponse pour lřarabe. Ce systèmeintègre des procédures de raisonnement automatique (RA) et des techniques dereconnaissance dřimplications textuelles (RTE). Pour le faire, nous nous appuyons sur desreprésentations sémantiques et logiques de la question et des passages. Ainsi, nousdéterminons lřimplication entre ces représentations logiques afin de trouver la réponseprécise
Design of a Hand Pose Recognition System for Mobile and Embedded Devices
No full textToday, smart devices such smart watches and smart cell phones are becoming ever-present in all fields that influence the quality of life of the modern people. These on-board systems have revolutionized the behavior of human beings and especially their way of communicating. In this context and to improve the experience of using these devices, we aim to develop a system that recognizes hand poses in the air by a smart device. In this work, the system is based on Histogram of Oriented Gradient (HOG) features and Support Vector Machine (SVM) classifier. The impact of using HOG and SVM on mobile devices is studied. To carry out this study, we used an improved version of the "NUS I" dataset and obtained a recognition rate of approximately 94%. In addition, we conducted run speed experiments on various mobile devices to study the impact of this task on this embedded platform. The main contribution of this work is to test the impact of using the HOG descriptor and the SVM classifier in terms of recognition rate and execution time on low-end smartphones.Today, smart devices such smart watches and smart cell phones are becoming ever-present in all fields that influence the quality of life of the modern people. These on-board systems have revolutionized the behavior of human beings and especially their way of communicating. In this context and to improve the experience of using these devices, we aim to develop a system that recognizes hand poses in the air by a smart device. In this work, the system is based on Histogram of Oriented Gradient (HOG) features and Support Vector Machine (SVM) classifier. The impact of using HOG and SVM on mobile devices is studied. To carry out this study, we used an improved version of the "NUS I" dataset and obtained a recognition rate of approximately 94%. In addition, we conducted run speed experiments on various mobile devices to study the impact of this task on this embedded platform. The main contribution of this work is to test the impact of using the HOG descriptor and the SVM classifier in terms of recognition rate and execution time on low-end smartphones
Towards a Statistical Approach to the Analysis, the Indexing, and the Semantic Search of Medical Videoconferences
No full textIn this article, the authors introduce their OSSVIRI information retrieval system which composed of three modules. In the analysis module, they have proposed a statistical technique exploiting the word frequency in order to extract the simple, compound and specific terms from the documents. In the indexing module, the authors used the ontology to associate the terms with their concepts, retrieve the relations between them and disambiguate the concepts to improve the sematic content of the documents. The concepts and relations are represented as a conceptual graph. In the research module, the authors have proposed a technique of users' query reformulation based on external resources and users' profiles and a technique of pairing based on the combined expansion of the requests and the documents guided by the context of the requirement in information and the documentary contents. This system is validated using the metrics from the research information and comparisons with existing statistical approach. The authors show that their approach achieves good results.</jats:p
Un système de récupération et de classification d’images extraites des caméras de vidéo-surveillance
No full textIn this thesis, we present a robust descriptor for background subtraction based on an unsupervised anomaly detection algorithm, called DeepSphere which is able to detect moving objects from video sequences. Unlike conventional background-foreground separation algorithms, this descriptor is less sensitive to noise and detects foreground objects without additional image processing. In addition, our proposal exploits both deep autoencoders and hy-persphere learning methods, having the ability to capture spatio-temporal dependencies between components and through "timesteps", to flexibly learn a non-linear feature representation and reconstruct normal behaviors from potentially anomalous input data. The high quality non-linear representations learned by the autoencoder helps the hypersphere to better distinguish anomalous cases by learning a compact boundary separating normal and ano-malous data. By adapting this algorithm to the background subtraction task, foreground objects are well captured by DeepSphere and the quality of detection of these objects is improved. Once these objects are detected (people/ cars ...), an approach is proposed to classify them using a DCGAN discriminator network in a semi-supervised manner. The discriminator is transformed into a multi-class classifier that uses both a large number of unlabeled data and a very small number of labeled data to compensate the lack of data and the high cost of collecting additional data or labeling all the data. Finally, we have adopted an approach based on FaceNet model to recognize the extracted people through their faces. In addition, we extended our proposal with a data augmentation method based on DCGANs instead of using standard data augmentation methods. This not only increases the accuracy of the model, but also reduces the execution time and the deep neural network learning time by almost half.Dans cette thèse, nous présentons un descripteur robuste pour la soustraction d’arrière-plan fondé sur un algorithme de détection des anomalies non-supervisé, appelé DeepSphere, capable de détecter les objets en mouvement dans les séquences vidéos. Contrairement aux algorithmes de séparation arrière-avant plan conventionnels, ce descripteur est tolérant aux variations d’illumination, robuste face aux bruits et aux régions d’arrière-plan dynamiques et détecte les objets de premier-plan sans utiliser de traitement d’image supplémentaire. En outre, ce descripteur exploite à la fois les autoencodeurs profonds et les méthodes d’apprentissage en hypersphère, ayant la capacité de capturer les dépendances spatio-temporelles entre les composants et à travers les pas de temps, d’apprendre de manière flexible une représentation non-linéaire des caractéristiques et de reconstruire les comportements normaux à partir des données d’entrée potentiellement anormales. Les représentations non linéaires de haute qualité apprises par l’autoencodeur aident l’hypersphère à mieux distinguer les cas anormaux en apprenant une frontière compacte séparant les données normales et anormales. En adaptant cet algorithme à la tâche de soustraction d’arrière-plan, les objets de premier plan sont bien capturés par DeepSphere et la qualité de la détection de ces objets est améliorée. Une fois que ces objets sont détectés (personnes/voitures...), une approche est proposée pour les classer en utilisant le réseau discriminateur du DCGAN de manière semi-supervisée. Le discriminateur est transformé en un classificateur multi-classes qui utilise à la fois un grand nombre de données non étiquetées et un très petit nombre de données étiquetées pour compenser la limite de manque de données et le coût élevé de collecte des données supplémentaires ou d’étiquetage de toutes les données. Enfin, nous avons proposé une approche basée sur le modèle FaceNet pour la reconnaissance faciale des personnes extraites. De plus, nous avons étendu notre proposition par une méthode d’augmentation des données basée sur DCGANs au lieu d’utiliser les méthodes standard d’augmentation des données. Cela augmente non seulement la précision du modèle, mais réduit aussi de près de moitié le temps d’exécution et le temps d’apprentissage du réseau neuronal profond
Integration of Multi-criteria Decision Analysis in the Evaluation Process of Interactive Adaptive Systems based on ISO/IEC 25040 Standard
No full textDuring the evaluation process of Interactive Adaptive Systems (IAS), there are some tasks that require making decisions in favor of the appropriate solution(s) out of several decision options that can be available. Numerous influencing factors should be considered surrounding these tasks. The aim of this thesis consists in supporting decision-making during the evaluation process of interactive adaptive systems, based on the international standard ISO/IEC 25040. In this way, a multi-criteria decision analysis approach and a Web-based multi-criteria decision support system are proposed in order to support a decision aid for a range of different possible stakeholders involved in the evaluation process. Firstly, the priorities of evaluation (sub)attributes to be assessed are identified. Then, the appropriate evaluation method(s) for the layered evaluation and the evaluation as a whole are proposed. Finally, we have investigated the application of our proposal through a case study following this evaluation processLors du processus d'évaluation des Systèmes Interactifs Adaptatifs (SIA), il existe certaines tâches qui nécessitent d'effectuer un choix d'une ou plusieurs solutions appropriées. Divers facteurs peuvent influencer ces décisions. L'objectif de cette thèse consiste à aider différents intervenants impliqués dans le processus d'évaluation à prendre des décisions appropriées en se basant sur la norme ISO/IEC 25040. Dans ce cadre, nous proposons une approche d'aide à la décision multicritère, ainsi qu'un système d'aide à la décision multicritère basé sur le Web. D’abord, nous nous intéressons à l’identification des niveaux de priorité des (sous)attributs d'évaluation qui doivent être évalués dans les SIA. Par la suite, nous nous focalisons sur la sélection des méthodes appropriées pour l'évaluation structurée et l'évaluation traditionnelle. De plus, nous proposons un processus d'évaluation adapté aux SIA, basé sur la norme ISO/IEC 25040. Enfin, un système interactif adaptatif dans le domaine du transport est étudié afin de valider notre proposition en se basant sur ce processus d'évaluation
Towards a Statistical Approach to the Analysis, the Indexing, and the Semantic Search of Medical Videoconferences
No full textIn this article, the authors introduce their OSSVIRI information retrieval system which composed of three modules. In the analysis module, they have proposed a statistical technique exploiting the word frequency in order to extract the simple, compound and specific terms from the documents. In the indexing module, the authors used the ontology to associate the terms with their concepts, retrieve the relations between them and disambiguate the concepts to improve the sematic content of the documents. The concepts and relations are represented as a conceptual graph. In the research module, the authors have proposed a technique of users' query reformulation based on external resources and users' profiles and a technique of pairing based on the combined expansion of the requests and the documents guided by the context of the requirement in information and the documentary contents. This system is validated using the metrics from the research information and comparisons with existing statistical approach. The authors show that their approach achieves good results.</jats:p
- …
