Dissertations / Theses on the topic 'Deep Learning, Database'

Traffic accidents are of concern for traffic safety; 1.25 million deaths are reported each year. Hence, it is crucial to have access to real-time data and rapidly detect or predict accidents. Predicting the occurrence of a highway car accident accurately any significant length of time into the future is not feasible since the vast majority of crashes occur due to unpredictable human negligence and/or error. However, rapid traffic incident detection could reduce incident-related congestion and secondary crashes, alleviate the waste of vehicles’ fuel and passengers’ time, and provide appropriate information for emergency response and field operation. While the focus of most previously proposed techniques is predicting the number of accidents in a certain region, the problem of predicting the accident occurrence or fast detection of the accident has been little studied. To address this gap, we propose a deep learning approach and build a deep neural network model based on long short term memory (LSTM). We apply it to forecast the expected speed values on freeways’ links and identify the anomalies as potential accident occurrences. Several detailed features such as weather, traffic speed, and traffic flow of upstream and downstream points are extracted from big datasets. We assess the proposed approach on a traffic dataset from Sacramento, California. The experimental results demonstrate the potential of the proposed approach in identifying the anomalies in speed value and matching them with accidents in the same area. We show that this approach can handle a high rate of rapid accident detection and be implemented in real-time travelers’ information or emergency management systems.
M.S.
Rapid traffic accident detection/prediction is essential for scaling down non-recurrent conges- tion caused by traffic accidents, avoiding secondary accidents, and accelerating emergency system responses. In this study, we propose a framework that uses large-scale historical traffic speed and traffic flow data along with the relevant weather information to obtain robust traffic patterns. The predicted traffic patterns can be coupled with the real traffic data to detect anomalous behavior that often results in traffic incidents in the roadways. Our framework consists of two major steps. First, we estimate the speed values of traffic at each point based on the historical speed and flow values of locations before and after each point on the roadway. Second, we compare the estimated values with the actual ones and introduce the ones that are significantly different as an anomaly. The anomaly points are the potential points and times that an accident occurs and causes a change in the normal behavior of the roadways. Our study shows the potential of the approach in detecting the accidents while exhibiting promising performance in detecting the accident occurrence at a time close to the actual time of occurrence.

L’identification de dispositifs dits sensibles est soumise à différentes contraintes de sécurité ou de consommation d’énergie, ce qui rend les méthodes d’identification classiques peu adaptées. Pour répondre à ces contraintes, il est possible d’utiliser les défauts intrinsèques de la chaîne de transmission des dispositifs pour les identifier. Ces défauts altèrent le signal transmis et créent alors une signature par nature unique et non reproductible appelée empreinte Radio Fréquence (RF). Pour identifier un dispositif grâce à son empreinte RF, il est possible d’utiliser des méthodes d’estimation d’imperfections pour extraire une signature qui peut être utilisée par un classifieur, ou bien d’utiliser des méthodes d’apprentissage telles que les réseaux de neurones. Toutefois, la capacité d’un réseau de neurones à reconnaître les dispositifs dans un contexte particulier dépend fortement de la base de données d’entraînement. Dans cette thèse, nous proposons un générateur de bases de données virtuelles basé sur des modèles de transmission et d’imperfections RF. Ces bases de données virtuelles permettent de mieux comprendre les tenants et aboutissants de l’identification RF et de proposer des solutions pour rendre l’identification plus robuste. Dans un second temps, nous nous intéressons aux problématiques de complexité de la solution d’identification via deux axes. Le premier consiste à utiliser des graphes programmables intriqués, qui sont des modèles d’apprentissage par renforcement, basés sur des techniques d’évolution génétique moins complexes que les réseaux de neurones. Le second axe propose l’utilisation de l’élagage sur des réseaux de neurones de la littérature pour réduire la complexité de ces derniers
Identifying so-called sensitive devices is subject to various security or energy consumption constraints, making conventional identification methods unsuitable. To meet these constraints, it is possible to use intrinsic faults in the device’s transmission chain to identify them. These faults alter the transmitted signal, creating an inherently unique and non-reproducible signature known as the Radio Frequency (RF) fingerprint. To identify a device using its RF fingerprint, it is possible to use imperfection estimation methods to extract a signature that can be used by a classifier, or to use learning methods such as neural networks. However, the ability of a neural network to recognize devices in a particular context is highly dependent on the training database. This thesis proposes a virtual database generator based on RF transmission and imperfection models. These virtual databases allow us to better understand the ins and outs of RF identification and to propose solutions to make identification more robust. Secondly, we are looking at the complexity of the identification solution in two ways. The first involves the use of intricate programmable graphs, which are reinforcement learning models based on genetic evolution techniques that are less complex than neural networks. The second is to use pruning on neural networks found in the literature to reduce their complexity

The alarm system plays a vital role to grant safety and reliability in the process industry. Ideally, an alarm should inform the operator about critical conditions only; during alarm floods, the operator may be overwhelmed by several alarms in a short time span. Crucial alarms are more likely to be missed during these situations. Poor alarm management is one of the main causes of unintended plant shut down, incidents and near misses in the chemical industry. Most of the alarms triggered during a flood episode are nuisance alarms –i.e. alarms that do not communicate new information to the operator, or alarms that do not require an operator action. Chattering alarms –i.e. that repeat three or more times in a minute, and redundant alarms –i.e. duplicated alarms, are common forms of nuisance. Identifying nuisance alarms is a key step to improve the performance of the alarm system. Advanced techniques for alarm rationalization have been developed, proposing methods to quantify chattering, redundancy and correlation between alarms. Although very effective, these techniques produce static results. Machine Learning appears to be an interesting opportunity to retrieve further knowledge and support these techniques. This knowledge can be used to produce more flexible and dynamic models, as well as to predict alarm behaviour during floods. The aim of this study is to develop a machine learning-based algorithm for real-time alarm classification and rationalization, whose results can be used to support the operator decision-making procedure. Specifically, efforts have been directed towards chattering prediction during alarm floods. Advanced techniques for chattering, redundancy and correlation assessment have been performed on a real industrial alarm database. A modified approach has been developed to dynamically assess chattering, and the results have been used to train three different machine learning models, whose performance has been evaluated and discussed.

Healthcare analytics has traditionally relied upon linear and logistic regression models to address clinical research questions mostly because they produce highly interpretable results [1, 2]. These results contain valuable statistics such as p-values, coefficients, and odds ratios that provide healthcare professionals with knowledge about the significance of each covariate and exposure for predicting the outcome of interest [1]. Thus, they are often favored over new deep learning models that are generally more accurate but less interpretable and scalable. However, the statistical power of linear and logistic regression is contingent upon satisfying modeling assumptions, which usually requires altering or transforming the data, thereby hindering interpretability. Thus, generalized additive models are useful for overcoming this limitation while still preserving interpretability and accuracy. The major research question in this work involves investigating whether particular sedative agents (fentanyl, propofol, versed, ativan, and precedex) are associated with different discharge dispositions for patients with acute traumatic brain injury (TBI). To address this, we compare the effectiveness of various models (traditional linear regression (LR), generalized additive models (GAMs), and deep learning) in providing guidance for sedative choice. We evaluated the performance of each model using metrics for accuracy, interpretability, scalability, and generalizability. Our results show that the new deep learning models were the most accurate while the traditional LR and GAM models ii i maintained better interpretability and scalability. The GAMs provided enhanced interpretability through pairwise interaction heat maps and generalized well to other domains and class distributions since they do not require satisfying the modeling assumptions used in LR. By evaluating the model results, we found that versed was associated with better discharge dispositions while ativan was associated with worse discharge dispositions. We also identified other significant covariates including age, the Northeast region, the Acute Physiology and Chronic Health Evaluation (APACHE) score, Glasgow Coma Scale (GCS), and ethanol level. The versatility of versed may account for its association with better discharge dispositions while ativan may have negative effects when used to facilitate intubation. Additionally, most of the significant covariates pertain to the clinical state of the patient (APACHE, GCS, etc.) whereas most non-significant covariates were demographic (gender, ethnicity, etc.). Though we found that deep learning slightly improved over LR and generalized additive models after fine-tuning the hyperparameters, the deep learning results were less interpretable and therefore not ideal for making the aforementioned clinical insights. However deep learning may be preferable in cases with greater complexity and more data, particularly in situations where interpretability is not as critical. Further research is necessary to validate our findings, investigate alternative modeling approaches, and examine other outcomes and exposures of interest.
Master of Science
Patients with Traumatic Brain Injury (TBI) often require sedative agents to facilitate intubation and prevent further brain injury by reducing anxiety and decreasing level of consciousness. It is important for clinicians to choose the sedative that is most conducive to optimizing patient outcomes. Hence, the purpose of our research is to provide guidance to aid this decision. Additionally, we compare different modeling approaches to provide insights into their relative strengths and weaknesses. To achieve this goal, we investigated whether the exposure of particular sedatives (fentanyl, propofol, versed, ativan, and precedex) was associated with different hospital discharge locations for patients with TBI. From best to worst, these discharge locations are home, rehabilitation, nursing home, remains hospitalized, and death. Our results show that versed was associated with better discharge locations and ativan was associated with worse discharge locations. The fact that versed is often used for alternative purposes may account for its association with better discharge locations. Further research is necessary to further investigate this and the possible negative effects of using ativan to facilitate intubation. We also found that other variables that influence discharge disposition are age, the Northeast region, and other variables pertaining to the clinical state of the patient (severity of illness metrics, etc.). By comparing the different modeling approaches, we found that the new deep learning methods were difficult to interpret but provided a slight improvement in performance after optimization. Traditional methods such as linear ii i regression allowed us to interpret the model output and make the aforementioned clinical insights. However, generalized additive models (GAMs) are often more practical because they can better accommodate other class distributions and domains.

Descrizione del processo di raccolta dati relativi all'inquinamento atmosferico ed alle condizioni meteorologiche in Emilia-Romagna. Introduzione alle principali tecniche di machine learning: le reti neurali artificiali. Utilizzo di alcuni framework specifici in tale ambito (TensorFlow, Keras) per la definizione di un modello capace di prevedere la concentrazione di un particolare inquinante (biossido di azoto), partendo dai dati raccolti nella prima fase. Descrizione e analisi dei risultati ottenuti.

In questa tesi analizzeremo le prestazioni di un System on Chip (SoC) in alcuni contesti di calcolo parallelo: in questo caso il SoC in questione è il Raspberry Pi. L'ultima versione rilasciata (Raspberry Pi 3 B+) è dotata di 1 GB di RAM e di un processore quad-core con architettura ARM NEON, caratterizzata dai prezzi e consumi ridotti. Tali specifiche danno i presupposti per un inserimento di questo dispositivo in contesti di high performace computing (HPC) tramite l'utilizzo di una programmazione parallela specifica. Per valutare le prestazioni di questa scheda si è voluto implementare un'applicazione che sia di uso comune in ambienti HPC, evitando semplici benchmark sulle singole componenti hardware. L'applicazione deve inoltre essere implementata tenendo conto dell'architettura di cui si dispone, in modo da ottenere risultati quanto più possibili legati alle caratteristiche hardware. Il progetto prevede di sviluppare un sistema di machine learning: una rete neurale artificiale che si addestra nel riconoscere cifre scritte a mano libera. Verranno quindi descritte e implementate alcune tecniche per parallelizzare la fase di addestramento della rete neurale. Questa applicazione verrà sfruttata per effettuare dei benchmark sia sul Raspberry Pi, che su un calcolatore "standard": l'università di Bologna ha messo a disposizione un server di calcolo che dispone di due processori Intel Xeon. I dati raccolti su queste due architetture (Raspberry Pi e Xeon) saranno messi a confronto. L'obiettivo è quello di analizzare se effettivamente dispositivi come il Raspberry Pi possono avere un qualche tipo di vantaggio in contesti HPC: in particolare verrà svolta un'analisi sulla capacità di svolgere calcolo parallelo.

Thermals are regions of rising hot air formed on the ground through the warming of the surface by the sun. Thermals are commonly used by birds and glider pilots to extend flight duration, increase cross-country distance, and conserve energy. This kind of powerless flight using natural sources of lift is called soaring. Once a thermal is encountered, the pilot flies in circles to keep within the thermal, so gaining altitude before flying off to the next thermal and towards the destination. A single thermal can net a pilot thousands of feet of elevation gain, however estimating thermal locations is not an easy task. Pilots look for different indicators: color variation on the ground because the difference in the amount of heat absorbed by the ground varies based on the color/composition, birds circling in an area gaining lift, and certain types of cloud formations (cumulus clouds). The above methods are not always reliable enough and pilots study the weather for thermals by estimating solar heating of the ground using cloud cover and time of year and the lapse rate and dew point of the troposphere. In this paper, we present a Machine Learning based solution for assisting in forecasting thermals. We created a custom dataset using flight data recorded and uploaded to public databases by soaring pilots. We determine where and when the pilot encountered thermals to pull weather and satellite images corresponding to the location and time of the flight. Using this dataset we train an algorithm to automatically predict the location of thermals given as input the current weather conditions and terrain information obtained from Google Earth Engine and thermal regions encountered as truth labels. We were able to converge very well on the training and validation set, proving our method with around a 0.98 F1 score. These results indicate success in creating a custom dataset and a powerful neural network with the necessity of bolstering our custom dataset.

La biométrie est de plus en plus utilisée à des fins d’identification compte tenu de la relation étroite entre la personne et son identifiant (comme une empreinte digitale). Nous positionnons cette thèse sur la problématique de l’identification d’individus à partir de ses empreintes digitales. L’empreinte digitale est une donnée biométrique largement utilisée pour son efficacité, sa simplicité et son coût d’acquisition modeste. Les algorithmes de comparaison d’empreintes digitales sont matures et permettent d’obtenir en moins de 500 ms un score de similarité entre un gabarit de référence (stocké sur un passeport électronique ou une base de données) et un gabarit acquis. Cependant, il devient très important de déterminer l'identité d'un individu contre une population entière en un temps très court (quelques secondes). Ceci représente un enjeu important compte tenu de la taille de la base de données biométriques (contenant un ensemble d’individus de l’ordre d’un pays). Par exemple, avant de délivrer un nouveau passeport à un individu qui en fait la demande, il faut faire une recherche d'identification sur la base des données biométriques du pays afin de s'assurer que ce dernier n'en possède pas déjà un autre mais avec les mêmes empreintes digitales (éviter les doublons). Ainsi, la première partie du sujet de cette thèse concerne l’identification des individus en utilisant les empreintes digitales. D’une façon générale, les systèmes biométriques ont pour rôle d’assurer les tâches de vérification (comparaison 1-1) et d’identification (1-N). Notre sujet se concentre sur l’identification avec N étant à l’échelle du million et représentant la population d’un pays par exemple. Dans le cadre de nos travaux, nous avons fait un état de l’art sur les méthodes d’indexation et de classification des bases de données d’empreintes digitales. Nous avons privilégié les représentations binaires des empreintes digitales pour indexation. Tout d’abord, nous avons réalisé une étude bibliographique et rédigé un support sur l’état de l’art des techniques d’indexation pour la classification des empreintes digitales. Ensuite, nous avons explorer les différentes représentations des empreintes digitales, puis réaliser une prise en main et l’évaluation des outils disponibles à l’imprimerie Nationale (IN Groupe) servant à l'extraction des descripteurs représentant une empreinte digitale. En partant de ces outils de l’IN, nous avons implémenté quatre méthodes d’identification sélectionnées dans l’état de l’art. Une étude comparative ainsi que des améliorations ont été proposées sur ces méthodes. Nous avons aussi proposé une nouvelle solution d'indexation d'empreinte digitale pour réaliser la tâche d’identification qui améliore les résultats existant. Les différents résultats sont validés sur des bases de données de tailles moyennes publiques et nous utilisons le logiciel Sfinge pour réaliser le passage à l’échelle et la validation complète des stratégies d’indexation. Un deuxième aspect de cette thèse concerne la sécurité. Une personne peut avoir en effet, la volonté de dissimuler son identité et donc de mettre tout en œuvre pour faire échouer l’identification. Dans cette optique, un individu peut fournir une empreinte de mauvaise qualité (portion de l’empreinte digitale, faible contraste en appuyant peu sur le capteur…) ou fournir une empreinte digitale altérée (empreinte volontairement abîmée, suppression de l’empreinte avec de l’acide, scarification…). Il s'agit donc dans la deuxième partie de cette thèse de détecter les doigts morts et les faux doigts (silicone, impression 3D, empreinte latente) utilisés par des personnes mal intentionnées pour attaquer le système. Nous avons proposé une nouvelle solution de détection d'attaque basée sur l'utilisation de descripteurs statistiques sur l'empreinte digitale. Aussi, nous avons aussi mis en place trois chaînes de détections des faux doigts utilisant les techniques d'apprentissages profonds
Biometrics are increasingly used for identification purposes due to the close relationship between the person and their identifier (such as fingerprint). We focus this thesis on the issue of identifying individuals from their fingerprints. The fingerprint is a biometric data widely used for its efficiency, simplicity and low cost of acquisition. The fingerprint comparison algorithms are mature and it is possible to obtain in less than 500 ms a similarity score between a reference template (enrolled on an electronic passport or database) and an acquired template. However, it becomes very important to check the identity of an individual against an entire population in a very short time (a few seconds). This is an important issue due to the size of the biometric database (containing a set of individuals of the order of a country). Thus, the first part of the subject of this thesis concerns the identification of individuals using fingerprints. Our topic focuses on the identification with N being at the scale of a million and representing the population of a country for example. Then, we use classification and indexing methods to structure the biometric database and speed up the identification process. We have implemented four identification methods selected from the state of the art. A comparative study and improvements were proposed on these methods. We also proposed a new fingerprint indexing solution to perform the identification task which improves existing results. A second aspect of this thesis concerns security. A person may want to conceal their identity and therefore do everything possible to defeat the identification. With this in mind, an individual may provide a poor quality fingerprint (fingerprint portion, low contrast by lightly pressing the sensor...) or provide an altered fingerprint (impression intentionally damaged, removal of the impression with acid, scarification...). It is therefore in the second part of this thesis to detect dead fingers and spoof fingers (silicone, 3D fingerprint, latent fingerprint) used by malicious people to attack the system. In general, these methods use machine learning techniques and deep learning. Secondly, we proposed a new presentation attack detection solution based on the use of statistical descriptors on the fingerprint. Thirdly, we have also build three presentation attacks detection workflow for fake fingerprint using deep learning. Among these three deep solutions implemented, two come from the state of the art; then the third an improvement that we propose. Our solutions are tested on the LivDet competition databases for presentation attack detection

Les chercheurs ont établi une forte corrélation entre les étés chauds et la fréquence ainsi que l'intensité desincendies de forêt. Le réchauffement climatique dû aux gaz à effet de serre tels que le dioxyde de carboneaugmente la température dans certaines parties du monde. Or, les incendies libèrent des quantitésimportantes de gaz à effet de serre, engendrant une augmentation de la température moyenne sur terreinduisant à son tour une augmentation des incendies de forêt... Les incendies détruisent des millionsd'hectares de zones forestières, des écosystèmes abritant de nombreuses espèces et ont un cout importantpour nos sociétés. La prévention et les moyens de maîtrise des incendies doivent être une priorité pour arrêtercette spirale infernale.Dans ce cadre, la détection de la fumée est très importante, car elle est le premier indice d'un début d'incendie.Le feu et surtout la fumée sont des objets difficiles à détecter dans les images visibles en raison de leurcomplexité en termes de forme, de couleur et de texture. Cependant, l'apprentissage profond couplé à lasurveillance vidéo peut atteindre cet objectif. L'architecture des réseaux de neurones convolutifs (CNN) estcapable de détecter avec une très bonne précision la fumée et le feu dans les images RVB. De plus, cesstructures peuvent segmenter la fumée ainsi que le feu en temps réel. La richesse de la base de donnéesd'apprentissage des réseaux profonds est un élément très important permettant une bonne généralisation.Ce manuscrit présente différentes architectures profondes basées sur des réseaux convolutifs permettant dedétecter et localiser la fumée et le feu dans les images vidéo dans le domaine du visible
Researchers have found a strong correlation between hot summers and the frequency and intensity of forestfires. Global warming due to greenhouse gases such as carbon dioxide is increasing the temperature in someparts of the world. Fires release large amounts of greenhouse gases, causing an increase in the earth'saverage temperature, which in turn causes an increase in forest fires... Fires destroy millions of hectares offorest areas, ecosystems sheltering numerous species and have a significant cost for our societies. Theprevention and control of fires must be a priority to stop this infernal spiral.In this context, smoke detection is very important because it is the first clue of an incipient fire. Fire andespecially smoke are difficult objects to detect in visible images due to their complexity in terms of shape, colorand texture. However, deep learning coupled with video surveillance can achieve this goal. Convolutionalneural network (CNN) architecture is able to detect smoke and fire in RGB images with very good accuracy.Moreover, these structures can segment smoke as well as fire in real time. The richness of the deep networklearning database is a very important element allowing a good generalization.This manuscript presents different deep architectures based on convolutional networks to detect and localizesmoke and fire in video images in the visible domain

Cette thèse de doctorat explore l’application des algorithmes d’apprentissage profond (Deep Learning - DL) pour relever les défis hydrologiques dans le bassin de la Seine, le deuxième plus long fleuve de France. Le régime hydraulique complexe de la Seine, influencé par les précipitations variables, les affluents, les interventions humaines et les fluctuations des marées, constitue un cadre idéal pour l’utilisation de techniques computationnelles avancées. Les modèles DL, notamment les réseaux neuronaux récurrents et les mécanismes d’attention, ont été choisis pour leur capacité à capturer les dépendances temporelles à long terme dans les séries chronologiques, surpassant ainsi les modèles traditionnels d’apprentissage machine (ML), tout en nécessitant moins de calibration manuelle que les modèles à base physique. La recherche se concentre sur le développement de méthodologies personnalisées pour améliorer l’efficacité du DL et optimiser son application face aux défis spécifiques du bassin de la Seine. Ces défis incluent la compréhension et la gestion des interactions complexes dans la zone d’étude, la gestion des limitations des données, le comblement des lacunes de données, la reconstruction et la projection des bases de données historiques manquantes, cruciales pour analyser les fluctuations du niveau d’eau en réponse à des variables telles que les changements climatiques, ainsi que la prédiction des événements d’inondations, notamment les crues extrêmes.Nos contributions, détaillées à travers quatre articles, illustrent l’efficacité du DL dans divers défis hydrologiques et applications : le comblement des lacunes des données de niveau d’eau sur plusieurs mois dans les enregistrements horaires, la reconstruction historique des paramètres de qualité de l’eau sur une période de plus de 15 ans dans le passé, l’analyse des interactions entre les stations, ainsi que la prédiction des événements de crue extrême à grande échelle (jusqu’à 7 jours à l’avance dans les données quotidiennes) et à petite échelle (jusqu’à 24 heures dans les données horaires).Les techniques proposées, telles que l’approche de décomposition Mini-Look-Back, les stratégies automatisées de reconstruction historique, les fonctions de perte personnalisées et l’ingénierie des caractéristiques approfondie, mettent en lumière la polyvalence et l’efficacité des modèles DL pour surmonter les limitations des données et surpasser les méthodes traditionnelles. La recherche souligne l’importance de l’interprétabilité en plus de la précision de la prédiction, offrant ainsi des informations sur la dynamique complexe de la Seine.Ces résultats mettent en évidence les potentialités du DL et des méthodologies développées dans les applications hydrologiques, tout en suggérant une applicabilité plus large à travers divers domaines traitant des séries chronologiques de données
This PhD thesis explores the application of deep learning (DL) algorithms to address hydrological challenges in the Seine River basin, France’s second longest river. The Seine’s intricate hydraulic regime, shaped by variable rainfall, tributaries, human interventions, and tidal fluctuations, presents an ideal scenario for advanced computational techniques. DL models, particularly recurrent-based neural networks and attention mechanisms, were chosen for their ability to capture long-term temporal dependencies in time series data, outperforming traditional machine learning (ML) models and their reduced need for manual calibration compared to physical-based models.The research focuses on developing custom methodologies to enhance DL efficiency and optimize its application to specific challenges within the Seine River Basin. Key challenges include addressing complex interactions within the study area, predicting extreme flood events, managing data limitations, and reconstructing missing historical databases crucial for analyzing water level fluctuations in response to variables such as climatic changes. The objective is to uncover insights, bridge data gaps, and enhance flood prediction accuracy, particularly for extreme events, thereby advancing smarter water management solutions.Detailed across four articles, our contributions showcase the effectiveness of DL in various hydrological challenges and applications: filling missing water level data gaps that may span several months in hourly records, projecting water quality parameters over 15 years in the past, analyzing station interactions, and predicting extreme flood events on both large (up to 7 days ahead in daily data) and small scales (up to 24 hours in hourly data).Proposed techniques such as the Mini-Look-Back decomposition approach, automated historical reconstruction strategies, custom loss functions, and extensive feature engineering highlight the versatility and efficacy of DL models in overcoming data limitations and outperforming traditional methods. The research emphasizes interpretability alongside prediction accuracy, providing insights into the complex dynamics of hydrological systems. These findings underscore the potential of DL and the developed methodologies in hydrological applications while suggesting broader applicability across various fields dealing with time series data

L’analyse d’images médicales joue un rôle essentiel en cardiologie pour la réalisation du diagnostique cardiaque clinique et le suivi de l’état du patient. Parmi les modalités d’imagerie utilisées, l’imagerie par ultrasons, temps réelle, moins coûteuse et portable au chevet du patient, est de nos jours la plus courante. Malheureusement, l’étape nécessaire de segmentation sémantique (soit l’identification et la délimitation précise) des structures cardiaques est difficile en échocardiographie à cause de la faible qualité des images ultrasonores, caractérisées en particulier par l’absence d’interfaces nettes entre les différents tissus. Pour combler le manque d’information, les méthodes les plus performante, avant ces travaux, reposaient sur l’intégration d’informations a priori sur la forme ou le mouvement du cœur, ce qui en échange réduisait leur adaptabilité au cas par cas. De plus, de telles approches nécessitent pour être efficaces l’identification manuelle de plusieurs repères dans l’image, ce qui rend le processus de segmentation difficilement reproductible. Dans cette thèse, nous proposons plusieurs algorithmes originaux et entièrement automatiques pour la segmentation sémantique d’images échocardiographiques. Ces méthodes génériques sont adaptées à la segmentation échocardiographique par apprentissage supervisé, c’est-à-dire que la résolution du problème est construite automatiquement à partir de données pré- analysées par des cardiologues entraînés. Grâce au développement d’une base de données et d’une plateforme d’évaluation dédiées au projet, nous montrons le fort potentiel clinique des méthodes automatiques d’apprentissage supervisé, et en particulier d’apprentissage profond, ainsi que la possibilité d’améliorer leur robustesse en intégrant une étape de détection automatique des régions d’intérêt dans l’image
The analysis of medical images plays a critical role in cardiology. Ultrasound imaging, as a real-time, low cost and bed side applicable modality, is nowadays the most commonly used image modality to monitor patient status and perform clinical cardiac diagnosis. However, the semantic segmentation (i.e the accurate delineation and identification) of heart structures is a difficult task due to the low quality of ultrasound images, characterized in particular by the lack of clear boundaries. To compensate for missing information, the best performing methods before this thesis relied on the integration of prior information on cardiac shape or motion, which in turns reduced the adaptability of the corresponding methods. Furthermore, such approaches require man- ual identifications of key points to be adapted to a given image, which makes the full process difficult to reproduce. In this thesis, we propose several original fully-automatic algorithms for the semantic segmentation of echocardiographic images based on supervised learning ap- proaches, where the resolution of the problem is automatically set up using data previously analyzed by trained cardiologists. From the design of a dedicated dataset and evaluation platform, we prove in this project the clinical applicability of fully-automatic supervised learning methods, in particular deep learning methods, as well as the possibility to improve the robustness by incorporating in the full process the prior automatic detection of regions of interest

The appearance of the world varies dramatically not only from place to place but also from hour to hour and month to month. Billions of images that capture this complex relationship are uploaded to social-media websites every day and often are associated with precise time and location metadata. This rich source of data can be beneficial to improve our understanding of the globe. In this work, we propose a general framework that uses these publicly available images for constructing dense maps of different ground-level attributes from overhead imagery. In particular, we use well-defined probabilistic models and a weakly-supervised, multi-task training strategy to provide an estimate of the expected visual and auditory ground-level attributes consisting of the type of scenes, objects, and sounds a person can experience at a location. Through a large-scale evaluation on real data, we show that our learned models can be used for applications including mapping, image localization, image retrieval, and metadata verification.

De nos jours, les systèmes de vision sont de plus en plus utilisés dans le contexte routier. Ils permettent ainsi d'assurer la sécurité et faciliter la mobilité. Ces systèmes de vision sont généralement affectés par la dégradation des conditions météorologiques en présence de brouillard ou de pluie forte, phénomènes limitant la visibilité et réduisant ainsi la qualité des images. Afin d'optimiser les performances des systèmes de vision, il est nécessaire de disposer d'un système de détection fiable de ces conditions météorologiques défavorables.Il existe des capteurs météorologiques dédiés à la mesure physique, mais ils sont coûteux. Ce problème peut être résolu en utilisant les caméras qui sont déjà installées sur les routes. Ces dernières peuvent remplir simultanément deux fonctions : l'acquisition d'images pour les applications de surveillance et la mesure physique des conditions météorologiques au lieu des capteurs dédiés. Suite au grand succès des réseaux de neurones convolutifs (CNN) dans la classification et la reconnaissance d'images, nous avons utilisé une méthode d'apprentissage profond pour étudier le problème de la classification météorologique. L'objectif de notre étude est de chercher dans un premier temps à mettre au point un classifieur du temps, qui permet de discriminer entre temps « normal », brouillard et pluie. Dans un deuxième temps, une fois la classe connue, nous cherchons à développer un modèle de mesure de la distance de visibilité météorologique du brouillard. Rappelons que l'utilisation des CNN exige l'utilisation de bases de données d'apprentissage et de test. Pour cela, deux bases de données ont été utilisées, "Cerema-AWP database" (https://ceremadlcfmds.wixsite.com/cerema-databases), et la base "Cerema-AWH database", en cours d'acquisition depuis 2017 sur le site de la Fageole sur l'autoroute A75. Chaque image des deux bases est labellisée automatiquement grâce aux données météorologiques relevées sur le site permettant de caractériser diverses gammes de pluie et de brouillard. La base Cerema-AWH, qui a été mise en place dans le cadre de nos travaux, contient cinq sous-bases : conditions normales de jour, brouillard fort, brouillard faible, pluie forte et pluie faible. Les intensités de pluie varient de 0 mm/h à 70 mm/h et les visibilités météorologiques de brouillard varient entre 50m et 1800m. Parmi les réseaux de neurones connus et qui ont montré leur performance dans le domaine de la reconnaissance et la classification, nous pouvons citer LeNet, ResNet-152, Inception-v4 et DenseNet-121. Nous avons appliqué ces réseaux dans notre système de classification des conditions météorologiques dégradées. En premier lieu, une étude justificative de l'usage des réseaux de neurones convolutifs est effectuée. Elle étudie la nature de la donnée d'entrée et les hyperparamètres optimaux qu'il faut utiliser pour aboutir aux meilleurs résultats. Ensuite, une analyse des différentes composantes d'un réseau de neurones est menée en construisant une architecture instrumentale de réseau de neurones. La classification des conditions météorologiques avec les réseaux de neurones profonds a atteint un score de 83% pour une classification de cinq classes et 99% pour une classification de trois classes.Ensuite, une analyse sur les données d'entrée et de sortie a été faite permettant d'étudier l'impact du changement de scènes et celui du nombre de données d'entrée et du nombre de classes météorologiques sur le résultat de classification.Enfin, une méthode de transfert de bases de données a été appliquée. Cette méthode permet d'étudier la portabilité du système de classification des conditions météorologiques d'un site à un autre. Un score de classification de 63% a été obtenu en faisant un transfert entre une base publique et la base Cerema-AWH. (...)
Nowadays, vision systems are becoming more and more used in the road context. They ensure safety and facilitate mobility. These vision systems are generally affected by the degradation of weather conditions, like heavy fog or strong rain, phenomena limiting the visibility and thus reducing the quality of the images. In order to optimize the performance of the vision systems, it is necessary to have a reliable detection system for these adverse weather conditions.There are meteorological sensors dedicated to physical measurement, but they are expensive. Since cameras are already installed on the road, they can simultaneously perform two functions: image acquisition for surveillance applications and physical measurement of weather conditions instead of dedicated sensors. Following the great success of convolutional neural networks (CNN) in classification and image recognition, we used a deep learning method to study the problem of meteorological classification. The objective of our study is to first seek to develop a classifier of time, which discriminates between "normal" conditions, fog and rain. In a second step, once the class is known, we seek to develop a model for measuring meteorological visibility.The use of CNN requires the use of train and test databases. For this, two databases were used, "Cerema-AWP database" (https://ceremadlcfmds.wixsite.com/cerema-databases), and the "Cerema-AWH database", which has been acquired since 2017 on the Fageole site on the highway A75. Each image of the two bases is labeled automatically thanks to meteorological data collected on the site to characterize various levels of precipitation for rain and fog.The Cerema-AWH base, which was set up as part of our work, contains 5 sub-bases: normal day conditions, heavy fog, light fog, heavy rain and light rain. Rainfall intensities range from 0 mm/h to 70mm/h and fog weather visibilities range from 50m to 1800m. Among the known neural networks that have demonstrated their performance in the field of recognition and classification, we can cite LeNet, ResNet-152, Inception-v4 and DenseNet-121. We have applied these networks in our adverse weather classification system. We start by the study of the use of convolutional neural networks. The nature of the input data and the optimal hyper-parameters that must be used to achieve the best results. An analysis of the different components of a neural network is done by constructing an instrumental neural network architecture. The conclusions drawn from this analysis show that we must use deep neural networks. This type of network is able to classify five meteorological classes of Cerema-AWH base with a classification score of 83% and three meteorological classes with a score of 99%Then, an analysis of the input and output data was made to study the impact of scenes change, the input's data and the meteorological classes number on the classification result.Finally, a database transfer method is developed. We study the portability from one site to another of our adverse weather conditions classification system. A classification score of 63% by making a transfer between a public database and Cerema-AWH database is obtained.After the classification, the second step of our study is to measure the meteorological visibility of the fog. For this, we use a neural network that generates continuous values. Two fog variants were tested: light and heavy fog combined and heavy fog (road fog) only. The evaluation of the result is done using a correlation coefficient R² between the real values and the predicted values. We compare this coefficient with the correlation coefficient between the two sensors used to measure the weather visibility on site. Among the results obtained and more specifically for road fog, the correlation coefficient reaches a value of 0.74 which is close to the physical sensors value (0.76)

Le sujet de la thèse est l'extraction et la segmentation des vêtements à partir d'images en utilisant des techniques de la vision par ordinateur, de l'apprentissage par ordinateur et de la description d'image, pour la recommandation de manière non intrusive aux utilisateurs des produits similaires provenant d'une base de données de vente. Nous proposons tout d'abord un extracteur d'objets dédié à la segmentation de la robe en combinant les informations locales avec un apprentissage préalable. Un détecteur de personne localises des sites dans l'image qui est probable de contenir l'objet. Ensuite, un processus d'apprentissage intra-image en deux étapes est est développé pour séparer les pixels de l'objet de fond. L'objet est finalement segmenté en utilisant un algorithme de contour actif qui prend en compte la segmentation précédente et injecte des connaissances spécifiques sur la courbure locale dans la fonction énergie. Nous proposons ensuite un nouveau framework pour l'extraction des vêtements généraux en utilisant une procédure d'ajustement globale et locale à trois étapes. Un ensemble de modèles initialises un processus d'extraction d'objet par un alignement global du modèle, suivi d'une recherche locale en minimisant une mesure de l'inadéquation par rapport aux limites potentielles dans le voisinage. Les résultats fournis par chaque modèle sont agrégés, mesuré par un critère d'ajustement globale, pour choisir la segmentation finale. Dans notre dernier travail, nous étendons la sortie d'un réseau de neurones Fully Convolutional Network pour inférer le contexte à partir d'unités locales (superpixels). Pour ce faire, nous optimisons une fonction énergie, qui combine la structure à grande échelle de l'image avec le local structure superpixels, en recherchant dans l'espace de toutes les possibilité d'étiquetage. De plus, nous introduisons une nouvelle base de données RichPicture, constituée de 1000 images pour l'extraction de vêtements à partir d'images de mode. Les méthodes sont validées sur la base de données publiques et se comparent favorablement aux autres méthodes selon toutes les mesures de performance considérées
The topic of the thesis is the extraction and segmentation of clothing items from still images using techniques from computer vision, machine learning and image description, in view of suggesting non intrusively to the users similar items from a database of retail products. We firstly propose a dedicated object extractor for dress segmentation by combining local information with a prior learning. A person detector is applied to localize sites in the image that are likely to contain the object. Then, an intra-image two-stage learning process is developed to roughly separate foreground pixels from the background. Finally, the object is finely segmented by employing an active contour algorithm that takes into account the previous segmentation and injects specific knowledge about local curvature in the energy function.We then propose a new framework for extracting general deformable clothing items by using a three stage global-local fitting procedure. A set of template initiates an object extraction process by a global alignment of the model, followed by a local search minimizing a measure of the misfit with respect to the potential boundaries in the neighborhood. The results provided by each template are aggregated, with a global fitting criterion, to obtain the final segmentation.In our latest work, we extend the output of a Fully Convolution Neural Network to infer context from local units(superpixels). To achieve this we optimize an energy function,that combines the large scale structure of the image with the locallow-level visual descriptions of superpixels, over the space of all possiblepixel labellings. In addition, we introduce a novel dataset called RichPicture, consisting of 1000 images for clothing extraction from fashion images.The methods are validated on the public database and compares favorably to the other methods according to all the performance measures considered

Recently, strategies of National Basketball Association teams have evolved with the skillsets of players and the emergence of advanced analytics. One of the most effective actions in dynamic offensive strategies in basketball is the dribble hand-off (DHO). This thesis proposes an architecture for a classification pipeline for detecting DHOs in an accurate and automated manner. This pipeline consists of a combination of player tracking data and event labels, a rule set to identify candidate actions, manually reviewing game recordings to label the candidates, and embedding player trajectories into hexbin cell paths before passing the completed training set to the classification models. This resulting training set is examined using the information gain from extracted and engineered features and the effectiveness of various machine learning algorithms. Finally, we provide a comprehensive accuracy evaluation of the classification models to compare various machine learning algorithms and highlight their subtle differences in this problem domain.

Le sujet de la thèse est l'extraction et la segmentation des vêtements à partir d'images en utilisant des techniques de la vision par ordinateur, de l'apprentissage par ordinateur et de la description d'image, pour la recommandation de manière non intrusive aux utilisateurs des produits similaires provenant d'une base de données de vente. Nous proposons tout d'abord un extracteur d'objets dédié à la segmentation de la robe en combinant les informations locales avec un apprentissage préalable. Un détecteur de personne localises des sites dans l'image qui est probable de contenir l'objet. Ensuite, un processus d'apprentissage intra-image en deux étapes est est développé pour séparer les pixels de l'objet de fond. L'objet est finalement segmenté en utilisant un algorithme de contour actif qui prend en compte la segmentation précédente et injecte des connaissances spécifiques sur la courbure locale dans la fonction énergie. Nous proposons ensuite un nouveau framework pour l'extraction des vêtements généraux en utilisant une procédure d'ajustement globale et locale à trois étapes. Un ensemble de modèles initialises un processus d'extraction d'objet par un alignement global du modèle, suivi d'une recherche locale en minimisant une mesure de l'inadéquation par rapport aux limites potentielles dans le voisinage. Les résultats fournis par chaque modèle sont agrégés, mesuré par un critère d'ajustement globale, pour choisir la segmentation finale. Dans notre dernier travail, nous étendons la sortie d'un réseau de neurones Fully Convolutional Network pour inférer le contexte à partir d'unités locales (superpixels). Pour ce faire, nous optimisons une fonction énergie, qui combine la structure à grande échelle de l'image avec le local structure superpixels, en recherchant dans l'espace de toutes les possibilité d'étiquetage. De plus, nous introduisons une nouvelle base de données RichPicture, constituée de 1000 images pour l'extraction de vêtements à partir d'images de mode. Les méthodes sont validées sur la base de données publiques et se comparent favorablement aux autres méthodes selon toutes les mesures de performance considérées
The topic of the thesis is the extraction and segmentation of clothing items from still images using techniques from computer vision, machine learning and image description, in view of suggesting non intrusively to the users similar items from a database of retail products. We firstly propose a dedicated object extractor for dress segmentation by combining local information with a prior learning. A person detector is applied to localize sites in the image that are likely to contain the object. Then, an intra-image two-stage learning process is developed to roughly separate foreground pixels from the background. Finally, the object is finely segmented by employing an active contour algorithm that takes into account the previous segmentation and injects specific knowledge about local curvature in the energy function.We then propose a new framework for extracting general deformable clothing items by using a three stage global-local fitting procedure. A set of template initiates an object extraction process by a global alignment of the model, followed by a local search minimizing a measure of the misfit with respect to the potential boundaries in the neighborhood. The results provided by each template are aggregated, with a global fitting criterion, to obtain the final segmentation.In our latest work, we extend the output of a Fully Convolution Neural Network to infer context from local units(superpixels). To achieve this we optimize an energy function,that combines the large scale structure of the image with the locallow-level visual descriptions of superpixels, over the space of all possiblepixel labellings. In addition, we introduce a novel dataset called RichPicture, consisting of 1000 images for clothing extraction from fashion images.The methods are validated on the public database and compares favorably to the other methods according to all the performance measures considered

碩士
國立中正大學
資訊工程研究所
105
In recent years, international exchanges and trade cooperation become frequent. Legal practice among nations increases as well. To ensure consistency of legal document, constructing a standard translation dictionary as a translation reference is necessary. However, there are plenty of laws and regulations which cover a wide range of area. If we manually mark all legal keywords, it will be highly inefficient. In this thesis, we propose an automatic Chinese legal keyword extraction algorithm based on deep learning technology, specifically, Back Propagation Neural Network (BPNN). This system helps extract legal keywords from legal documents which legal expert would identify. This system consists of two parts: candidate keyword generation and keyword identification. First, the keyword candidate set is generated by using the word segmentation and combination method proposed in this study. Compared with word segmentation without combination, this method effectively improves the coverage of the actual legal keyword set. Furthermore, compared with the traditional n-gram combination of Chinese words, it can significantly reduce the number of keyword candidates and cost less time on following classification. To identify legal keywords based on BPNN, specific features are first defined based on the characteristics of legal keywords identified by experts to improve the performance of BPNN. A real world data set is then used to train the BPNN. Experimental results show the effectiveness of the proposed approach with the overall average accuracy, precision, recall and f-measure values of 92.6%, 89.2%, 88.2% and 88.2% respectively. All of these measures have been significantly improved as compared to our previous work.

Yes
The most recent face recognition systems are mainly dependent on feature representations obtained using either local handcrafted-descriptors, such as local binary patterns (LBP), or use a deep learning approach, such as deep belief network (DBN). However, the former usually suffers from the wide variations in face images, while the latter usually discards the local facial features, which are proven to be important for face recognition. In this paper, a novel framework based on merging the advantages of the local handcrafted feature descriptors with the DBN is proposed to address the face recognition problem in unconstrained conditions. Firstly, a novel multimodal local feature extraction approach based on merging the advantages of the Curvelet transform with Fractal dimension is proposed and termed the Curvelet–Fractal approach. The main motivation of this approach is that theCurvelet transform, a newanisotropic and multidirectional transform, can efficiently represent themain structure of the face (e.g., edges and curves), while the Fractal dimension is one of the most powerful texture descriptors for face images. Secondly, a novel framework is proposed, termed the multimodal deep face recognition (MDFR)framework, to add feature representations by training aDBNon top of the local feature representations instead of the pixel intensity representations. We demonstrate that representations acquired by the proposed MDFR framework are complementary to those acquired by the Curvelet–Fractal approach. Finally, the performance of the proposed approaches has been evaluated by conducting a number of extensive experiments on four large-scale face datasets: the SDUMLA-HMT, FERET, CAS-PEAL-R1, and LFW databases. The results obtained from the proposed approaches outperform other state-of-the-art of approaches (e.g., LBP, DBN, WPCA) by achieving new state-of-the-art results on all the employed datasets.

碩士
國立中興大學
電機工程學系所
107
To solve the problem of getting lost indoors of an unfamiliar building, Our team proposes photos taken by a mobile phone and its inertia information to position and sketch the indoor floor plan. The deep learning algorithm in the research of this team requires a lot of indoor information in different rooms. It is not easy to obtain real-world data.Therefore, this paper uses the virtual data output by SketchUp as an aid. This paper focuses on providing virtual data by using SketchUp and Python automation. This thesis also trains a deep learning algorithm by different data sets (virtual cat/dog and real cat/dog). The result shows Transfer Learning can improve the accuracy.