HAL Portal IOGS (nstitut d'Optique Graduate School)
Not a member yet
12589 research outputs found
Sort by
Study of radioluminescent silica optical fiber for quality control in proton therapy
No full textInternational audienc
Reconnaissance des émotions au sein d'images riches en contexte : exploration d'architectures compactes orientées robotique sociale
No full textWe interact daily with colleagues, relatives, or sometimes even strangers. Given the growing interest in social robotics, it would not be surprising if in the years to come it would become just as common to converse with robots capable of assisting or entertaining us. As human beings, we know how to adapt our behavior according to the reactions of our interlocutors: for example, we know when to speak, when to listen, when to intervene, or how to choose our words. Ultimately, it is these adaptive abilities that machines still lack so that they are truly capable of offering more natural interactions. In order to improve human-robot exchanges, it is therefore necessary to fully exploit the verbal and non-verbal information emitted by human beings, such as their emotions, and to integrate them into the decision-making process of robots.The main non-verbal indicator of a person's emotional state is their facial expression. However, on images acquired at a relatively long distance, the face is not always fully exploitable. Fortunately, other clues can also help predict a person's emotion: their posture, the action they perform, the place represented, or even the visible objects. All these elements actually provide context, which helps extrapolate the emotion. In the literature, it is reffered as Context-Aware Emotion Recognition (CAER), and it is in this research area that the thesis work presented in this document falls.First, we developed an innovative approach based on deep learning, called BENet. Unlike state-of-the-art multi-stream methods that process people present in an image sequentially, our neural network analyzes them all simultaneously by generating emotion maps. In addition, BENet works directly from the raw image and is intended to be compact, since it is composed only of a common backbone on which specialized heads are grafted. We thus avoid the pre-processing steps, limit the necessary computing resources, and work in multi-tasks to promote its training and include the detection of people. After validating our approach on the data used in CAER, we implemented it on the ROS platform to make its integration possible within a robot.Following this initial work and still with the idea of remaining on compact but efficient models, we then explored a complementary solution to multi-task learning, namely knowledge amalgamation. This particular technique makes it possible to group together the knowledge of several teacher models, trained on specific data for various tasks, within a single student architecture. To do this, we project the representations extracted by all of these models into a common, automatically learned space, and force the student to reproduce a summarized version of those of the teachers. We were thus able to improve the performance of the student model, who thanks to our process surpasses his teachers in CAER without becoming more complex.Finally, we were interested in the distillation of multi-teacher knowledge in a simplified research framework: the Visual Emotion Analysis (VEA). We were thus able to highlight the complementarity of teacher models with diverse and varied pre-learned knowledge, and showed that it was possible to group them together taking into account their relevance within a single student model, thus greatly improving its performance. Also, despite its simplicity, the student model finally becomes capable of competing with the more complex approaches in the literature, or even beating them.Nous sommes quotidiennement amenés à interagir avec des collègues, des proches, ou parfois même avec des inconnus. Étant donné l'intérêt croissant porté à la robotique sociale, il ne serait pas surprenant que dans les années à venir il devienne tout aussi commun de converser avec des robots capables de nous assister ou de nous divertir. Nous savons en tant qu'êtres humains adapter notre comportement en fonction des réactions de nos interlocuteurs : nous savons par exemple quand prendre la parole, quand écouter, quand intervenir, ou comment choisir nos mots. Ce sont finalement ces capacités d'adaptation qui manquent encore aux machines pour qu'elles soient réellement capables de proposer des interactions plus naturelles. Afin d'améliorer les échanges humain-robot, il est donc nécessaire d'exploiter pleinement les informations verbales et non-verbales émises par les êtres humains, comme par exemple leurs émotions, et de les intégrer au processus de prise de décisions des robots.Le principal indicateur non-verbal de l’état émotionnel d’une personne est son expression faciale. Néanmoins, sur des images acquises à relativement longue distance, le visage n'est pas toujours pleinement exploitable. Heureusement, d'autres indices peuvent également permettre de prédire l'émotion d'une personne : sa posture, l'action qu'il effectue, le lieu représenté, ou encore les objets visibles. Tous ces éléments apportent en fait du contexte, qui aide à extrapoler l'émotion. Dans la littérature, il est alors question de Reconnaissance des Émotions à l'Aide du Contexte (REAC), et c'est dans ce cadre de recherche que s'inscrivent les travaux de thèse présentés dans ce document.Tout d'abord, nous avons développé une approche innovante basée sur l'apprentissage profond, appelée BENet. Contrairement aux méthodes multi-flux de l'état de l'art qui traitent les personnes présentes dans une image de façon séquentielle, notre réseau de neurones les analyse toutes simultanément en générant des cartes d'émotions. En outre, BENet travaille directement depuis l'image brute et se veut compact, puisqu'il n'est composé que d'une base commune sur laquelle viennent se greffer des têtes spécialisées. Nous évitons ainsi les étapes de pré-traitement, limitons les ressources de calcul nécessaires, et travaillons en multi-tâches pour favoriser son entraînement et inclure la détection des personnes. Après avoir validé notre approche sur les données utilisées en REAC, nous l'avons implémentée sur la plateforme ROS pour rendre son intégration possible au sein d'un robot.Suite à ces premiers travaux et toujours dans l'idée de rester sur des modèles compacts mais performants, nous avons alors exploré une solution complémentaire à l'apprentissage multi-tâches, à savoir l'amalgamation de connaissances. Cette technique particulière permet de regrouper les savoirs de plusieurs modèles enseignants, entraînés sur des données spécifiques pour des tâches variées, au sein d'une seule architecture élève. Pour cela, nous projetons les représentations extraites par l'ensemble de ces modèles dans un espace commun appris automatiquement, et forçons l'élève à reproduire une version résumée de celles des enseignants. Nous avons ainsi pu améliorer les performances de ce dernier, qui, grâce à notre processus, dépasse ses enseignants pour la REAC sans pour autant se complexifier.Pour finir, nous nous sommes intéressés à la distillation de connaissances multi-enseignants dans un cadre de recherche connexe : l'Analyse Visuelle des Émotions (AVE). Nous avons ainsi pu mettre en avant la complémentarité de modèles enseignants aux connaissances pré-apprises diverses et variées, et montré qu'il était possible de les regrouper en tenant compte de leur pertinence au sein d'un seul modèle élève, améliorant ainsi grandement ses performances. Aussi, malgré sa simplicité, le modèle élève devient finalement capable de rivaliser avec les approches plus complexes de la littérature, voire de les battre
Beyond color: in-situ acquisition of monumental ornaments
No full textInternational audienceAcquisition of the appearance on monumental ornaments is highly complex, due to their high dimensionality (geometry, lighting and observation directions, but also spatial variations in reflection properties). Although it is possible to obtain accurate, dense data in a controlled environment using calibrated devices, the task is far more complex in-situ.We summarize the main principles of the portable acquisition method developed by Corentin Cou during his Ph.D. [1]. It consists of simultaneously acquiring and reconstructing the object's shape, SVBRDF and normal maps, using two cameras, a light spot and two mirror spheres. We also propose a new method for calibrating photographic projector light sources. Our approach allows great freedom and simplicity in the choice of different light/view pairs. This approach has been validated on large-scale decorative elements for the restoration of Gaston de Saint-Maurice's residence in Cairo
Lightweight Active Fences for FPGAs
No full textInternational audienceThe use of active fences has been proposed as a protection against remote power analysis attacks. This counter-measure relies on reserving a reconfigurable space within the FPGA which will separate it into sub-regions. These “fences” will then generate some electrical interference to hinder the performance of an attack. As FPGAs can be configured in multiple ways, there are different approaches for connecting the hardware inside the fence. In this work, we describe a LUT-based configuration which can achieve the same instantaneous power drop as a ring oscillator bank with less LUTs. This contributes to reducing the hardware costs of active fences
CNN-based approach for 3D artifact correction of intensity diffraction tomography images
No full textInternational audience3D reconstructions after tomographic imaging often suffer from elongation artifacts due to the limited-angle acquisitions. Retrieving the original 3D shape is not an easy task, mainly due to the intrinsic morphological changes that biological objects undergo during their development. Here we present to the best of our knowledge a novel approach for correcting 3D artifacts after 3D reconstructions of intensity-only tomographic acquisitions. The method relies on a network architecture that combines a volumetric and a 3D finite object approach. The framework was applied to time-lapse images of a mouse preimplantation embryo developing from fertilization to the blastocyst stage, proving the correction of the axial elongation and the recovery of the spherical objects. This work paves the way for novel directions on a generalized non-supervised pipeline suited for different biological samples and imaging conditions
Solid-state Li metal battery with hybrid electrolyte: An overview of the Horizon Europe SEATBELT project
No full textInternational audienc
Multiscale characterization of the wettability of fs-laser textured thin film metallic glasses surfaces
No full textInternational audienceWith the absence of crystalline defects and their amorphous structure, metallic glasses (MGs) exhibit very interesting mechanical and chemical properties. They have been studied since the 60s in their bulk state (BMGs), but are size limited and complex to synthesize due to their high needed number of elements. More recently, PVD processes enabling high cooling rate of the deposited atoms have demonstrated the easier formation of metastable amorphous metallic phases, together with great freedom in the films’ chemistry.From pure metallic targets, the magnetron sputtering process has already shown its ability to synthesize binary Zr-Cu thin film metallic glasses (TFMGs) over a wide range of chemical compositions (from 13 to 85 at.% of Cu [1]), with low surface roughness together with the absence of grain boundaries, making them suitable for a femtosecond laser treatment [2], to further improve their properties.The work proposed here considers the formation of laser induced periodic surface structures (LIPSS) at the surface of two ternary magnetron sputtered TFMGs (ZrCuAg and ZrTiAg, with interesting biological properties [3]) using infrared ultrashort laser treatment. These textured surfaces are first studied in terms of topographic and chemical modifications, then a focus on the wettability modifications (hydro-phily/phoby) is proposed. Wettability is studied first at the macroscale from the conventional measurement of the water contact angle. On the other hand, the condensation process of water onto the surface is also measured at the microscale by in situ measurements conducted in an environmental scanning electron microscope (ESEM). Such a complementary small-scale method gives key information on the interaction of very small water droplets with the textured surface, opening the way to biological behavior of such surfaces.[1] M. Apreutesei, et al., “Zr-Cu thin film metallic glasses: An assessment of the thermal stability and phases transformation mechanisms”, Journal of Alloys and Compounds, 2015[2] M. Prudent, et al., “Initial morphology and feedback effects on laser-induced periodic nano-structuring of thin-film metallic glasses”, Nanomaterials, 2021[3] A. Etiemble, et al., “Innovative Zr-Cu-Ag thin film metallic glass deposited by magnetron PVD sputtering for antibacterial applications”, Journal of Alloys and Compounds, 201
Topographic, chemical and property modifications of PVD ZrCu-based thin film metallic glasses through femtosecond laser treatment
No full textInternational audienceMetallic glasses (MGs) exhibit very interesting mechanical and chemical properties due to their absence of crystalline defects and their amorphous structure. They have been studied a lot in their bulk state (BMGs), but are size limited and complex to synthesize. More recently, it was proved that PVD processes exhibit high cooling rate of the deposited atoms to allow the formation of metastable amorphous metallic phases [1]. Thus, metallic glasses are easier to obtain by PVD in thin film form than bulk ones. The synthesis of such thin films through PVD process allows a high freedom in the film chemistry composition.From pure metallic targets, the magnetron sputtering process has already shown its ability to synthesize Zr-based thin film metallic glasses (TFMGs). These films exhibited a very low surface roughness together with the absence of grain boundaries, making them suitable for a femtosecond laser treatment [2], in order to still improve their properties. The laser irradiation process modification with a great repeatability, and gives rise to localized topographic and chemical modifications at the surface of the thin film.The work proposed here considers the formation of laser induced periodic surface structures (LIPSS) at the surface of two ternary magnetron sputtered TFMGs (ZrCuAg and ZrTiAg) using femtosecond laser pulses treatment. The choice of the laser parameters leads to the formation of several surface texturations of the thin films. The characterization of these structures is made at several scales, from topographic to chemical modifications, from Atomic Force Microscopy to Electronic Microscopy (Scanning and Transmission EM). These measurements permit to better understand the interaction of ultrashort laser pulses to amorphous material like MGs. Then, some property modifications of the TFMGs induced by such laser textures are studied.Ref.:[1] C.-Y. Chuand, et al., “Mechanical properties study of a magnetron-sputtered Zr-based thin film metallic glass”, Surface and Coatings Technology, 2013[2] M. Prudent, et al., “Initial morphology and feedback effects on laser-induced periodic nano-structuring of thin-film metallic glasses”, Nanomaterials, 202
Ultrafast laser-induced topochemistry on metallic glass surfaces
No full textInternational audienceManufacturing multifunctional nanocomposite materials and engineered surface nanopatterns involves a strategic blend of topography, crystal structures, and chemistry. Here, we report the controllable formation of crystalline nanoparticles and intermetallic compounds on thin films of metallic glasses (Zr 50 Cu 50 , Ti 50 Cu 50 , and Zr 67 Ag 33 ) irradiated by ultrafast laser beams. Mapping the structural modification of the photoexcited and subsequently heated alloys reveals previously neglected chemical reactions with air, offering a direct solution for incorporating nanoparticles into an amorphous oxide matrix and broadening the range of laser-induced surface self-organization features. Our findings are attributed to the occurrence and enrichment of oxygen surface contamination that reacts with selected elements of the metallic glasses. Additionally, the growth of the crystalline phase from undercooled liquid may originate from the dissolution of oxides. Finally, our results establish that the combination of crystalline nanoparticles on amorphous periodic patterns can be universally obtained in a wide range of binary systems of irradiated metallic glasses.</p