La spectroscopie de transit est un outil puissant pour la caractérisation de l'atmosphère d'exoplanètes. Plusieurs phénomènes peuvent contaminer un spectre de transmission, dont l'hétérogénéité de la surface de l'étoile hôte due à l'activité stellaire. À basse résolution spectrale, la différence entre le cordon de transit et le reste de la surface y laisse des signatures qui pourraient être attribuées à tort à la planète. Les risques associés incluent des biais sur la mesure du rayon et des abondances atmosphériques de planètes. Afin de trouver une solution à ce problème, cette étude consiste à modéliser en détail des surfaces stellaires et des spectres de transit à basse et à très haute résolution. On cherche d'une part à qualifier l'ampleur du problème à basse résolution et, d'autre part, à déterminer si la haute résolution permet d'isoler la contamination stellaire et ainsi résoudre le problème. La modélisation se concentre sur trois types d'étoiles, entre K hâtive et M tardive. Les modèles confirment l'importance du problème et l'absence de solution évidente à basse résolution, principalement pour les étoiles M. À haute résolution, on parvient à séparer les signaux de la planète et de l'activité stellaire. Cela permet de briser l'ambiguïté à basse résolution, pourvu que la planète ait une variation de vitesse radiale suffisante durant le transit. Ces résultats soulignent la valeur d'un suivi à haute résolution lorsque possible. Même avec le télescope James-Webb, il sera difficile d'avoir totalement confiance en les résultats de caractérisation d'atmosphères utilisant des données à basse résolution. / Transit spectroscopy is a powerful tool for the characterisation of exoplanet atmospheres. There exist multiple sources of contamination for transmission spectra, including stellar activity induced surface heterogeneities on the host star. At low spectral resolution, differences between the transit chord and the rest of the surface leave signatures in the spectra, which could then be wrongly associated with the planet. This can introduce biases in radius and atmospheric abundance measurements of exoplanets. To solve this issue, this study consists in carefully modeling stellar surfaces and transit spectra at low and very high spectral resolution. We seek to, on one hand, understand the importance of the problem at low resolution, and, on the other hand, determine if high resolution allows us to isolate stellar contamination and thus solve this problem. Modeling is focused on three types of stars, from early K to late M. Models confirm the importance of the issue and the absence of an obvious solution at low resolution, especially for M stars. At high resolution, we manage to effectively split the planet and stellar activity signals. This allows us to break the ambiguity from low resolution, provided the planet experiences a sufficient radial velocity variation during transit. These results highlight the strong value of high resolution follow-ups when feasible. Even with the James-Webb space telescope, it will be difficult to fully trust the results of atmospheric abundance retrievals using low resolution data.

This project explores the understudied, yet important, case of human pose estimation in low-resolution images. This is done in the use-case of images with football players of known scale in the image. Human pose estimation can mainly be done in two different ways, the bottom-up method and the top-down method. This project explores the bottom-up method, which first finds body keypoints and then groups them to get the person, or persons, within the image. This method is generally faster and has been shown to have an advantage when there is occlusion or crowded scenes, but suffers from false positive errors. Low-resolution makes human pose estimation harder, due to the decreased information that can be extracted. Furthermore, the output heatmap risks becoming too small to correctly locate the keypoints. However, low-resolution human pose estimation is needed in many cases, if the camera has a low-resolution sensor or the person occupies a small portion of the image. Several neural networks are evaluated and, in conclusion, there are multiple ways to improve the current state of the art network HigherHRNet for lower resolution human pose estimation. Maintaining large feature maps through the network turns out to be crucial for low-resolution images and can be achieved by modifying the feature extractor in HigherHRNet. Furthermore, as the resolution decreases, the need for sub-pixel accuracy grows. To improve this, various heatmap encoding-decoding methods are investigated, and by using unbiased data processing, both heatmap encoding-decoding and coordinate system transformation can be improved. / Detta projekt utforskar det understuderade, men ändå viktiga, fallet med uppskattning av mänskliga poser i lågupplösta bilder. Detta görs i användningsområdet av bilder med fotbollsspelare av en förutbestämd storlek i bilden. Mänskliga poseuppskattningar kan huvudsakligen göras på två olika sätt, nedifrån-och-upp- metoden och uppifrån-och-ned-metoden. Detta projekt utforskar nedifrån-och- upp-metoden, som först hittar kroppsdelar och sedan grupperar dem för att få fram personen, eller personerna, i bilden. Denna metod är generellt sett snabbare och har visat sig vara fördelaktig i scenarion med ocklusion eller mycket folk, men lider av falska positiva felaktigheter. Låg upplösning gör uppskattning av mänskliga poser svårare, på grund av den minskade informationen som kan extraheras. Dessutom riskerar färgdiagramet att bli för liten för att korrekt lokalisera kroppsdelarna. Ändå behövs uppskattning av lågupplöst mänskliga poser i många fall, exempelvis om kameran har en lågupplöst sensor eller om personen upptar en liten del av bilden. Flera neurala nätverk utvärderas och sammanfattningsvis finns flera sätt att förbättra det nuvarande toppklassade nätverket HigherHRNet för uppskattning av mänskliga poser med lägre upplösning. Att bibehålla stora särdragskartor genom nätverket visar sig vara avgörande för lågupplösta bilder och kan uppnås genom att modifiera särdragsextraktorn i HigherHRNet. Dessutom, när upplösningen minskar, ökar behovet av subpixel-noggrannhet. För att förbättra detta undersöktes olika färgdiagram-kodning-avkodningsmetoder, och genom att använda opartisk databehandling kan både färgdiagram-kodning-avkodning och koordinatsystemtransformationen förbättras.

<div>Given the increasing importance of mobile data access, extending broadband wireless access have become a global grand challenge. Wireless sensor networks (WSNs) and millimeter wave (mmWave) systems have been introduced to resolve these issues which motivate us to have further investigation. In this paper, the first two work assuming a quantized-and-forward WSN. We first develop a rate adaptive integer forcing source coding (RAIF) scheme to enhance the system throughput by assigning optimal quantization rate to each sensor optimally. Then, we are interested in developing an supervised online technique for solving classification problems. In order to enhance the classification performance, we developed this technique by jointly training the decision function that determines/estimates class label, quantizers across all sensors, and reliability of sensors such that M' most reliable sensors are enabled. Finally, we develop an idea to provide a folded low-resolution ADC array architecture that can utilize any of the widely published centralized folded ADC (FADC) implementation by placing the centralized FADC branches at different antenna elements in a millimeter wave (mmWave) system. With adding a simple analog shift and modulo operations prior to the sign quantizer, we show that the multiple low-resolution ADCs across the array elements can be properly designed such that they can be combined into an effective high-resolution ADC with excellent performance characteristics.</div>

La classification des images revêt un intérêt majeur dans de nombreuses tâches de reconnaissance visuelle, en particulier pour la reconnaissance de véhicules au sol via les systèmes aéroportés, où les images traitées sont de faible résolution du fait de la large distance entre le porteur et la scène observée. Durant l'apprentissage, des données complémentaires peuvent être disponibles, qu'il s'agisse de connaissances sur les conditions de prise de vue ou de la version haute-résolution des images. Dans nos travaux, on s'intéresse au problème de la reconnaissance d'images faiblement résolues en prenant en compte des informations complémentaires pendant l'apprentissage. On montre d'abord l'intérêt des réseaux convolutionnels profonds pour la reconnaissance d'images faiblement résolues, en proposant notamment une architecture apprise sur les données. D'autre part, on s'appuie sur le cadre de l'apprentissage avec information privilégiée pour bénéficier des données d'entraînement complémentaires, ici les versions haute-résolution des images. Nous proposons deux méthodes d'intégration de l'information privilégiée dans l'apprentissage des réseaux de neurones. Notre premier modèle s'appuie sur ces données complémentaires pour calculer un niveau de difficulté absolue, attribuant un poids important aux images les plus facilement reconnaissables. Notre deuxième modèle introduit une contrainte de similitude entre les modèles appris sur chaque type de données. On valide expérimentalement nos deux modèles dans plusieurs cas d'application, notamment dans un contexte orienté grain-fin et sur une base de données contenant du bruit d'annotation. / Image classification has a prominent interest in numerous visual recognition tasks, particularly for vehicle recognition in airborne systems, where the images have a low resolution because of the large distance between the system and the observed scene. During the training phase, complementary data such as knowledge on the position of the system or high-resolution images may be available. In our work, we focus on the task of low-resolution image classification while taking into account supplementary information during the training phase. We first show the interest of deep convolutional networks for the low-resolution image recognition, especially by proposing an architecture learned on the targeted data. On the other hand, we rely on the framework of learning using privileged information to benefit from the complementary training data, here the high-resolution versions of the images. We propose two novel methods for integrating privileged information in the learning phase of neural networks. Our first model relies on these complementary data to compute an absolute difficulty level, assigning a large weight to the most easily recognized images. Our second model introduces a similarity constraint between the networks learned on each type of data. We experimentally validate our models on several application cases, especially in a fine-grained oriented context and on a dataset containing annotation noise.

No description available.

yes / PICARD is a space-based observatory hosting the Solar Diameter Imager and Surface Mapper (SODISM) telescope, which has continuously observed the Sun from July 2010 and up to March 2014. In order to study the fine structure of the solar surface, it is helpful to apply techniques that enhance the images so as to improve the visibility of solar features such as sunspots or faculae. The objective of this work is to develop an innovative technique to enhance the quality of the SODISM images in the five wavelengths monitored by the telescope at 215.0 nm, 393.37 nm, 535.7 nm, 607.1 nm and 782.2 nm. An enhancement technique using interpolation of the high-frequency sub-bands obtained by Discrete Wavelet Transforms (DWT) and the input image is applied to the SODISM images. The input images are decomposed by the DWT as well as Stationary Wavelet Transform (SWT) into four separate sub-bands in horizontal and vertical directions namely, low-low (LL), low-high (LH), high-low (HL) and high–high (HH) frequencies. The DWT high frequency sub-bands are interpolated by a factor 2. The estimated high frequency sub-bands (edges) are enhanced by introducing an intermediate stage using a stationary Wavelet Transform (SWT), and then all these sub-bands and input image are combined and interpolated with half of the interpolation factor α/2, used to interpolate the high-frequency sub-bands, in order to reach the required size for IDWT processing. Quantitative and visual results show the superiority of the proposed technique over a bicubic image resolution enhancement technique. In addition, filling factors for sunspots are calculated from SODISM images and results are presented in this work.

Pour mieux appréhender les mécanismes de transfert des gaz (CO2 et CH4) dissous dans l’eau porale des formations des marnes du site de Bure et des argiles à Opalinus (AOP) du Mt Terri dans leur état initial, un capteur infrarouge (IRTF) et un banc optique Raman innovant ont été développés et installés en laboratoire souterrain. Ces capteurs sont intégrés au dispositif de l’expérimentation « d’équilibration de gaz » développée par l’Andra (PAC) qui vise à suivre l’évolution d’une phase gazeuse initialement neutre au contact de la formation et de son eau porale grâce à un forage effectué et maintenu dans des conditions anaérobiques. Ces capteurs permettent un suivi quantitatif in situ et en ligne des gaz libérés par la formation à basse pression totale (<1,3 bar). Les modèles quantitatifs développés pour la mesure infrarouge ont une erreur relative moyenne de 1,66 % pour la pCO2 (mbar.m) et de 1,37 % pour la pCH4 (mbar.m). L’instrumentation IR d’un forage sur le site du Mt Terri et de deux forages sur le site de Bure (faciès C2b1 et C2d) a permis d’obtenir les courbes de transfert des deux gaz. Les courbes de transfert du CH4 ont été modélisées par un modèle de diffusion-advection qui ont permis l’évaluation de la concentration locale en CH4 dissous dans l’eau porale : elles sont comprises entre 3,06 et 14,23 mg.L-1 pour les AOP, et entre 0,36 et 1,28 mg.L-1 pour le faciès C2d et entre 0,56 et 1,55 mg.L-1 pour le faciès C2b1 des argilites de Bure. On montre que les équilibres eau/gaz/roche gouvernent la pCO2 après forage alors que la diffusion/advection explique son évolution sur le long terme. Une origine intraformationnelle des alcanes dissous est envisagée / An infrared sensor (IRTF) and an innovative Raman optical bench were implemented and developed in underground laboratories to improve our knowledge about migration mechanisms of dissolved gases (CO2 and CH4). This study is focussed on the characterisation of the initial state of the porewater of Callovo-Oxfordian marl (Bure) and Opalinus Clay (Mt Terri Middle Jurassic). These sensors are integrated into experimental devices of gas-equilibration test developed by Andra (PAC) to follow the gaseous phase behaviour in contact with the rock formation through a borehole drilled and maintained in anaerobic conditions, and initially filled with pure argon. These in situ sensors allow, on line quantitative analysis of gases released by the rock formation at low bulk pressure (<1.3 bar). Quantitative models were developed to transform peak intensities in partial pressures of gas. They give mean absolute relative errors about 1.66 % for pCO2 (mbar.m) and 1.37 % for pCH4 (mbar.m). Three years of IR monitoring of one borehole on the site of Mt Terri and two boreholes on the site of Bure (facies C2b1 and C2d) have been led. CH4 transfer curves were modelled by diffusion-advection. CH4 concentration in porewater from non-perturbed rock formation is estimated from all the experiments: concentrations between 3.06 and 14.23 mg.L-1 was obtained for Opalinus Clay, between 0.36 and 1.28 mg.L-1 for C2d facies and between 0.56 and 1.55 mg.L-1 for C2b1 facies in Callovo-Oxfordian marls of Bure. Gas/rock/water balance governs pCO2 after drilling, whereas diffusion/advection laws explain CO2 long-term profiles. An intra-formational origin of the organic gases can be proposed

Submitted by Marlene Santos (marlene.bc.ufg@gmail.com) on 2018-11-05T18:46:12Z No. of bitstreams: 2 Dissertação - Dyego Oliveira Almeida - 2018.pdf: 32110237 bytes, checksum: 4d30b2227be666802bbb2e5437450fc2 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Approved for entry into archive by Luciana Ferreira (lucgeral@gmail.com) on 2018-11-06T10:03:22Z (GMT) No. of bitstreams: 2 Dissertação - Dyego Oliveira Almeida - 2018.pdf: 32110237 bytes, checksum: 4d30b2227be666802bbb2e5437450fc2 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) / Made available in DSpace on 2018-11-06T10:03:22Z (GMT). No. of bitstreams: 2 Dissertação - Dyego Oliveira Almeida - 2018.pdf: 32110237 bytes, checksum: 4d30b2227be666802bbb2e5437450fc2 (MD5) license_rdf: 0 bytes, checksum: d41d8cd98f00b204e9800998ecf8427e (MD5) Previous issue date: 2018-10-03 / Coordenação de Aperfeiçoamento de Pessoal de Nível Superior - CAPES / The urban growth of the population and the deforestation of greenareas are one of the most critical problems currently in Brazil. Due to mobilization of rural people to the urban, high solar irradiation and the deforestation, the Government is creating sustainable actions sustainable in order to enlarge the green areas and permeable. In this perspective, to promote this mapping effectively in large areas necessary to the use of technologies of recognition of facial features. Low-resolution satellite imagery have low cost and great coverage area coverage, and therefore apply them in identifying features is advantageous over other types of images. However, to accomplish this identification is computationally complex due to the different features present in images of this type. This work proposes an effective method of digital processing of low resolution images in the identification of features, in particular the pertinent green aáreas with average accuracy of 80.5% and detection of buildings with an average accuracy of 63%. / O crescimento urbano da população e o desmatamento de áreas verdes são um dos problemas mais críticos atualmente no Brasil. Devido a mobilização da população rural para o âmbito urbano, elevação da irradiação solar e o desmatamento, o governo está criando ações sustentáveis a fim de ampliar as áreas verdes e permeáveis. Nesta perspectiva, para promover esse mapeamento de forma eficaz em grandes áreas se faz necessário o uso de tecnologias de reconhecimento de feições. Imagens de satélite de baixa resolução possuem baixo custo e grande abrangência de área, e portanto aplicá-las na identificação de feições é vantajoso em relação a outros tipos de imagens. No entanto, realizar essa identificação é computacionalmente complexo devido as diferentes características existentes em imagens desse tipo. Este trabalho propõe um método eficaz de processamento digital de imagens de baixa resolução na identificação de feições, em particular as pertinentes a áreas verdes com acurácia média de 80,5% e detecção de edificações com precisão média de 63%.

[pt] Os sistemas de múltiplas entradas e múltiplas saídas (MIMO) empregam um número crescente de antenas, o que leva a relevantes consumo de energia e custo de hardware dos front-ends correspondentes. Nesse contexto, o uso de conversores analógico-digitais (ADCs) de baixa resolução é promovido como uma solução promissora para este problema. Neste estudo consideramos um receptor MIMO de baixa resolução que implica que os sinais recebidos são processados simultaneamente pelos 1-bit ADCs e pela rede comparadora. Os sinais de entrada da rede comparadora podem vir de antenas diferentes, de modo que a extensão da rede comparadora pode ser interpretada como canais virtuais com saídas binárias. Com base nesses receptores MIMO de baixa resolução, desenvolvemos um estimador de canal e detector lineares de baixa resolução baseados no critério de mínimo erro médio quadrático (LRA-LMMSE) de acordo com o teorema de Bussgang. Duas redes de comparação são propostas, nomeadas, redes total e parcialmente conectadas. Também desenvolvemos uma rede parcialmente conectada baseada em busca gananciosa que usa muito menos comparadores para obter um desempenho bem próximo ao da rede totalmente conectada. Os resultados numéricos mostram que adicionar canais virtuais pode ser melhor do que adicionar canais físicos extras que correspondem a antenas de recepção adicionais em termos de taxa de erro de bit (BER). Além disso, ao empregar o estimador de canal proposto e seu erro de estimativa correspondente, construímos um limite inferior na taxa de soma ergódica para o receptor LRA-MMSE. Os resultados de simulação mostram que os sistemas com a proposta sistemas MIMO auxiliados por rede com quantização de 1-bit no receptor superam o convencional sistema MIMO de 1-bit em termos de desempenho de BER e erro quadrático médio (MSE). Além disso, as simulações numéricas confirmam uma vantagem significativa em termos de taxa de soma para o sistema proposto. / [en] Multiple-input multiple-output (MIMO) systems employs an increasing number of antennas, which leads to relevant energy consumption and hardware cost of the corresponding front ends. In this context, the use of lowresolution analog to digital converters (ADCs) is promoted as a promising solution to this problem. In this study we consider a low-resolution MIMO receiver which implies that the received signals simultaneously are processed by the 1-bit ADCs and the comparator network. The input signals for the comparator network can come from different antennas, such that the comparator network extension can be interpreted as virtual channels with binary outputs. Based on such low-resolution MIMO receivers, we develop low-resolution aware linear minimum mean-squared error (LRA-LMMSE) channel estimator and detector according to the Bussgang theorem. Two comparator networks are proposed, namely, fully and partially connected networks. We also devise a greedy search-based partially connected network that can use much less comparators to approach the performance of the fully connected network. Numerical results shows that adding virtual channels can be better than adding extra physical channels which corresponds to additional receive antennas in terms of bit error rate (BER). Furthermore, by employing the proposed channel estimator and its corresponding estimation error, we build up a lower bound on the ergodic sum rate for the LRA-LMMSE receiver. Simulation results show that the systems with the proposed network-aided MIMO systems with 1-bit quantization at the receiver outperforms the conventional 1-bit MIMO system in terms of BER and mean-square error (MSE) performances. Moreover, numerical simulations confirm a significant advantage in terms of sum rate for the proposed system.

Building occupancy estimation has become an important topic for sustainable buildings that has attracted more attention during the pandemics. Estimating building occupancy is a considerable problem in computer vision, while computer vision has achieved breakthroughs in recent years. But, machine learning algorithms for computer vision demand large datasets that may contain users’ private information to train reliable models. As privacy issues pose a severe challenge in the field of machine learning, this work aims to develop a privacypreserved machine learningbased method for people counting using a lowresolution thermal camera with 32 × 24 pixels. The method is applicable for counting people in different scenarios, concretely, counting people in spaces smaller than the field of view (FoV) of the camera, as well as large spaces over the FoV of the camera. In the first scenario, counting people in small spaces, we directly count people within the FoV of the camera by Multiple Object Detection (MOD) techniques. Our MOD method achieves up to 56.8% mean average precision (mAP). In the second scenario, we use Multiple Object Tracking (MOT) techniques to track people entering and exiting the space. We record the number of people who entered and exited, and then calculate the number of people based on the tracking results. The MOT method reaches 47.4% multiple object tracking accuracy (MOTA), 78.2% multiple object tracking precision (MOTP), and 59.6% identification F-Score (IDF1). Apart from the method, we create a novel thermal images dataset containing 1770 thermal images with proper annotation. / Uppskattning av hur många personer som vistas i en byggnad har blivit ett viktigt ämne för hållbara byggnader och har fått mer uppmärksamhet under pandemierna. Uppskattningen av byggnaders beläggning är ett stort problem inom datorseende, samtidigt som datorseende har fått ett genombrott under de senaste åren. Algoritmer för maskininlärning för datorseende kräver dock stora datamängder som kan innehålla användarnas privata information för att träna tillförlitliga modeller. Eftersom integritetsfrågor utgör en allvarlig utmaning inom maskininlärning syftar detta arbete till att utveckla en integritetsbevarande maskininlärningsbaserad metod för personräkning med hjälp av en värmekamera med låg upplösning med 32 x 24 pixlar. Metoden kan användas för att räkna människor i olika scenarier, dvs. att räkna människor i utrymmen som är mindre än kamerans FoV och i stora utrymmen som är större än kamerans FoV. I det första scenariot, att räkna människor i små utrymmen, räknar vi direkt människor inom kamerans FoV med MOD teknik. Vår MOD-metod uppnår upp till 56,8% av den totala procentuella fördelningen. I det andra scenariot använder vi MOT-teknik för att spåra personer som går in i och ut ur rummet. Vi registrerar antalet personer som går in och ut och beräknar sedan antalet personer utifrån spårningsresultaten. MOT-metoden ger 47,4% MOTA, 78,2% MOTP och 59,6% IDF1. Förutom metoden skapar vi ett nytt dataset för värmebilder som innehåller 1770 värmebilder med korrekt annotering.