International Professional University of Technology in Nagoya Repository
Not a member yet
15131 research outputs found
Sort by
An end-to-end learning solution for assessing the quality of Wikipedia articles
No full textInternational audienceWikipedia is considered as the largest knowledge repository in the history of humanity and plays a crucial role in modern daily life. Assigning the correct quality class to Wikipedia articles is an important task in order to provide guidance for both authors and readers of Wikipedia. Manual review cannot cope with the editing speed of Wikipedia. An automatic classification is required to classify quality of Wikipedia articles. Most existing approaches rely on traditional machine learning with manual feature engineering, which requires a lot of expertise and effort. Furthermore, it is known that there is no general perfect feature set, because information leak always occurs in feature extraction phase. Also, for each language of Wikipedia a new feature set is required. In this paper, we present an approach relying on deep learning for quality classification of Wikipedia articles. Our solution relies on Recurrent Neural Networks (RNN) which is an end-to-end learning technique that eliminates disadvantages of feature engineering. Our approach learns directly from raw data without human intervention and is language-neutral.Experimental results on English, French and Russian Wikipedia datasets show that our approach outperforms state-of-the-art solutions
Estimation de l'incertitude lors de la prévision de la production d'énergie d'origine photovoltaïque pour améliorer l'efficacité et la gestion d'un micro-réseau
No full textNational audienc
Parameter Design and Energy Control of the Power Train in a Hybrid Electric Boat
No full textInternational audienceWith the continuous development worldwide of the inland shipping industry, emissions to the atmosphere have become a serious threat in terms of pollution. Hybrid power technology is an important means for reducing pollution due to emissions from ships. This paper considers a power train series in a hybrid electric inland waterway boat. From the analysis of the structure and principle of the power train, the parameter design for its key devices is presented, and a novel energy control strategy is proposed. Navigation experience shows that the proposed design method and control strategy are useful and satisfactory
Learning-based Adaptive Tone Mapping for Keypoint Detection
No full textInternational audienceThe goal of tone mapping operators (TMOs) has traditionally beento display high dynamic range (HDR) pictures in a perceptually favorableway. However, when tone-mapped images are to be usedfor computer vision tasks such as keypoint detection, these designapproaches are suboptimal. In this paper, we propose a newlearning-based adaptive tone mapping framework which aims at enhancingkeypoint stability under drastic illumination variations. Tothis end, we design a pixel-wise adaptive TMO which is modulatedbased on a model derived by Support Vector Regression (SVR) usinglocal higher order characteristics. To circumvent the difficultyto train SVR in this context, we further propose a simple detection similarity-maximizationmodel to generate appropriate training samplesusing multiple images undergoing illumination transformations.We evaluate the performance of our proposed framework in termsof keypoint repeatability for state-of-the-art keypoint detectors. Experimentalresults show that our proposed learning-based adaptiveTMO yields higher keypoint stability when compared to existingperceptually-driven state-of-the-art TMOs
Hierarchical Dataflow Model for Efficient Programming of Clustered Manycore Processors
No full textInternational audienceProgramming Multiprocessor Systems-on-Chips (MPSoCs) with hundreds of heterogeneous Processing Elements (PEs), complex memory architectures, and Networks-on-Chips (NoCs) remains a challenge for embedded system designers. Dataflow Models of Computation (MoCs) are increasingly used for developing parallel applications as their high-level of abstraction eases the automation of mapping, task scheduling and memory allocation onto MPSoCs. This paper introduces a technique for deploying hierarchical dataflow graphs efficiently onto MPSoC. The proposed technique exploits different granularity of dataflow parallelism to generate both NoC-based communications and nested OpenMP loops. Deployment of an image processing application on a many-core MPSoC results in speedups of up to 58.7 compared to the sequential execution
Vers la transformation de la parole oesophagienne en voix laryngée à l'aide de techniques de conversion vocale
No full textNational audienceCe travail concerne le développement d'un système de conversion de voix oesophagienne dans le but est de rendre plus intelligible celle-ci. La conversion de voix est une technique de transformation d'un signal de parole d'un locuteur source, de manière à ce qu'il semble, à l'écoute, être prononcé par un locuteur cible. Etant donnée la spécificité de la voix oesophagienne, nous proposons dans cette étude d'appliquer une nouvelle technique de conversion vocale en tenant compte de la particularité de l'appareil vocal des patients qui ont subi une ablation de larynx. En effet, l'ablation des cordes vocales perturbe profondément le signal glottique et par conséquent la voix oesophagienne acquise par le patient laryngectomisé est difficile à comprendre, rauque et faible en intensité. Dans la littérature, plusieurs techniques de conversion des voix ont été proposées, parmi lesquelles, la technique du codage linéaire prédictif pour la conversion vocale [1] et la régression linéaire multi-variée [2] qui vise à réduire la discontinuité et la distorsion spectrale
Reconnaissance Statistique de la Parole Continue pour Voix Laryngée et Alaryngée
No full textLa Reconnaissance Automatique de la Parole (RAP) demeure depuis toujours un défiscientifique. Au cours de ces dernières années de grands efforts de recherche ont étéconcrétisés, afin de développer des systèmes d’aide et des solutions permettant d’effectuercertaine tâches jusqu’ici réservées aux humains. La parole est un mode de communicationnaturel, et un moyen facile pour échanger des informations entre humains. Unepersonne laryngectomisée, n’a pas la capacité de parler normalement puisqu’elle est dépourvuede ses cordes vocales suite à une ablation chirurgicale du larynx. Ainsi, le patientperd toute possibilité de communication avec une voix laryngée. Néanmoins, la rééducationavec un orthophoniste lui permet d’acquérir une voix de substitution dite “oesophagienne”.Contrairement à la parole laryngée (normale), cette parole oesophagienne(alaryngée) est rauque, faible en énergie et en intelligibilité ce qui la rend difficile à comprendre.L’objectif de cette thèse est la réalisation d’un système de reconnaissance automatiquede la parole oesophagienne (alaryngée). Ce système devrait être en mesure de restituer,la plus grande partie des informations phonétiques contenues dans le signal dela parole oesophagienne. Cette information textuelle fournie par la partie décodage de cesystème pourra être utilisée par un synthétiseur texte-parole (Text-To-Speech) dans le butde reconstruire une voix laryngée. Un tel système permettrait aux personnes laryngectomisées,une communication orale plus facile avec d’autres personnes.Notre première contribution est relative au développement d’un système de reconnaissanceautomatique de la parole laryngée en utilisant des modèles de Markov cachés.Les rares corpus de parole oesophagienne existants, ne sont pas dédiés à la reconnaissance,à cause d’un manque de données (souvent quelques dizaines de phrases sont enregistrées).Pour cette raison, nous avons conçu notre propre base de données dédiée àla reconnaissance de la parole oesophagienne contenant 480 phases prononcées par unlocuteur laryngectomisé.Dans une seconde partie, le système de reconnaissance de la parolelaryngée créé a été adapté et appliqué à cette parole oesophagienne. Notre dernièrecontribution au sujet de cette thèse concerne la réalisation d’un système hybride (correction= conversion + reconnaissance) fondé sur la conversion de la voix en projetant lesvecteurs acoustiques de la parole oesophagienne dans un espace moins perturbé et relatifà la parole laryngée. Nous montrons que ce système hybride est capable d’améliorer lareconnaissance de cette parole alaryngée.Automatic Speech Recognition (ASR) has always been a scientist challenge. Many researchefforts have been made over recent years to offer solutions and aiding systems inorder to carry out various tasks previously dedicated only to humans. Speech is consideredthe most natural mode of communication, and an easy way for exchanging informationbetween humans. A laryngectomee person lacks the ability of speaking normallybecause he/her lost his/her vocal cords after a surgical ablation of the larynx. Thus, thepatient loses the phonation ability. Only a reeducation by a speech therapist allows thisperson to provide a new substitution voice called “esophageal”. Unlike laryngeal speech(normal), esophageal speech (alaryngeal) is hoarse, weak in intensity and in intelligibilitywhichmakes it difficult to understand.The goal of this thesis is the implementation of an automatic esophageal speech (alaryngeal)recognition system. This system should be able to provide most of the phoneticinformation contained in the esophageal speech signal. The decoding part of this systemconnected to a text-to-speech synthesizer should allow the reconstruction of a laryngealvoice. Such a system should permit laryngectomees an easier oral communication withother people.Our first contribution concerns the development of an automatic laryngeal speech recognitionsystem using hidden Markov models. The few existing corpora of esophagealspeech, are not dedicated to recognition, because of a lack of data (only a few dozen sentencesare registered in practice). For this reason, we designed our own database dedicatedto esophageal speech recognition containing 480 sentences spoken by a laryngectomeespeaker. In the second part, our devoted laryngeal speech recognition system hasbeen adapted and applied to this esophageal speech. Our last contribution of this thesisconcerns the realization of a hybrid system (correction = conversion + recognition) basedon voice conversion by projecting the acoustic feature vectors of esophageal speech in aless disturbed space related to laryngeal speech. We demonstrate that this hybrid systemis able to improve the recognition of alaryngeal speech
Better Platooning Control Toward Autonomous Driving : An LTE Device-to-Device Communications Strategy That Meets Ultralow Latency Requirements
No full textInternational audiencePlatooning is the first step toward fully autonomous driving, which is deemed as one of the most representative fifth-generation (5G) use cases. Spacing and speed in a platoon of vehicles are regulated by a fully automated control system that relies on updated vehicles? kinematics data. In this article, we investigate the potential of long-term evolution (LTE) device-todevice (D2D) communications for data dissemination in the platoon. Exploiting pooled LTE resources and the coordination by the in-front vehicle of the platoon, the proposed solution is shown to fulfill the ultralow latency requirements of messaging in the platoon. As a further advantage, our proposal is able to provide spatial reuse of LTE resources among members of the same platoon and of different platoons, thus drastically reducing the capacity demand
A Novel Threshold based Compressed Channel Sensing in OFDM System
No full textInternational audienceHigh speed data transmission for wireless communication in orthogonal frequency division multiplexing (OFDM) system requires effective channel state information (CSI). CSI should be precisely estimated with low consumption of spectral resources and acceptable computational cost. To realize this goal, an effective compressed sensing (CS) based channel estimation scheme is proposed for sparse channels with large delay spreads, without prior knowledge of channel statistics and noise standard deviation. By fully considering the rank of the measurement matrix, a novel algorithm based on orthogonal matching pursuit (OMP) and least squares (LS) methods with a new threshold is proposed for effective channel estimation. Simulation results show that with fewer number of pilots, the proposed method outperforms the compared existing channel estimation methods in a comprehensive way and approaches the optimal channel estimation performance
On Idempotent Discrete Uninorms
No full textInternational audienceIn this paper we provide two axiomatizations of the class of idempotent discrete uninorms as conservative binary operations, where an operation is conservative if it always outputs one of its input values. More precisely we first show that the idempotent discrete uninorms are exactly those operations that are conservative, symmetric, and nondecreasing in each variable. Then we show that, in this characterization , symmetry can be replaced with both bisymmetry and existence of a neutral element