Teses / dissertações sobre o tema "Deeplearning"

Machine Learning e Deep Learning nell'ambito medico. Focus sull'imaging medico. Analisi di un DataSet con 188 radiografie di toraci tramite addestramento con Reti Neurali Convoluzionali, mostrando andamento del Misclassification Rate al variare dei parametri della Rete.

L’électrocardiogramme (ECG) est un outil non invasif permettant d’évaluer l’activité électrique du cœur. Ils sont largement utilisés dans la détection d’anomalies cardiaques. Les algorithmes d’apprentissage profond permettent la détection automatique de schémas complexes dans les données ECG, ce qui offre un potentiel important pour l’amélioration du diagnostic médical. Toutefois, leur adoption est freinée par un faible niveau de confiance des cliniciens et un besoin massif de données pour entrainer les modèles. L’intelligence artificielle, en particulier l’apprentissage profond (deep learning), permet d’explorer des représentations hiérarchiques de données complexes, ce qui permet de mieux comprendre les interactions internes. Néanmoins, l’interprétabilité des modèles est cruciale pour gagner la confiance des spécialistes et permettre une utilisation générale. Ces travaux de thèse, réalisés en étroite collaboration avec des spécialistes en cardiologie, visent à développer un nouvel algorithme d’interprétabilité pour les réseaux de neurones appliqués aux données ECG. Notre étude se concentre sur une pathologie cardiaque spécifique, la Torsades de pointes (TdP). La TdP est une arythmie mortelle associée à divers facteurs, notamment médicamenteux et/ou des mutations congénitales. Une prédiction précise de ce risque peut améliorer les soins aux patients et potentiellement sauver des vies. Nous avons commencé par concevoir un réseau de neuronnes pour prédire le risque de TdP à l’aide de données ECG. Ensuite, nous avons développé un nouvel algorithme d’interprétabilité baptisé Evocclusion, qui permet de mieux comprendre le processus de décision du réseau de neurones. Cet algorithme vise à fournir des informations lisibles par l’homme sur les prédictions du modèle, afin d’accroître la confiance des cliniciens et des spécialistes. Enfin, nous présentons deux autres méthodes développées pour améliorer l’analyse de l’ECG et la méthode d’interprétabilité. La qualité du signal est un aspect crucial dans l’analyse d’ECGs. Ainsi, nous proposons une nouvelle méthode utilisant un autoencodeur de débruitage pour réduire de manière significative le bruit présent dans les données ECG et reconstruire partiellement le signal. Cette technique améliore la fiabilité des données d’entrée pour des analyses approfondies et garantit que les réseaux de neurones ont accès à des informations de haute qualité. Nous avons également développé des réseaux supplémentaires pour segmenter l’ECG et extraire les battements, les ondes P et T et complexes QRS. Cette segmentation permet une compréhension plus approfondie des composants de l’ECG et ouvre la voie à de nouvelles analyses sur des composantes spécifiques du signal. En outre, nous fournissons une méthode pour évaluer un vecteur score de qualité ECG, ce qui nous permet de nous concentrer sur les parties du signal qui ont un bon score de qualité. Cette approche garantit que les informations les plus fiables sont utilisées pour l’analyse et les cliniciens, ce qui réduit le risque de faux positifs et négatifs. Cette recherche vise à renforcer la confiance dans l’utilisation de réseau de neurones, ce qui permettra d’améliorer l’automatisation des tâches complexes en médecine et ailleurs, et, en fin, d’améliorer le traitement des patients
Electrocardiograms (ECGs) are non-invasive tools for assessing the electrical activity of the heart, they are widely used to detect cardiac abnormalities. Deep learning algorithms enable automatic detection of complex patterns in ECG data, offering significant potential for improved cardiac diagnosis. However, their adoption is hindered by a low level of trust among medical professionals and a substantial need for data to train the models. Artificial intelligence, particularly deep learning, allows for exploration of hierarchical representations of complex data, leading to a better understanding of internal interactions. Nevertheless, interpretability of the models are crucial to gain specialists’ trust and facilitate widespread implementation. This thesis aims to develop a novel interpretability algorithm for neural networks applied to ECG analysis, working in close collaboration with cardiology specialists. Our study focuses on a specific cardiac pathology, Torsades-de-Pointes (TdP). TdP is a life threatening arrhythmia associated with various factors, including medications and congenital mutations. Accurate prediction of this risk can enhance patient care and potentially save lives. We started by designing a neural network algorithm for predicting the risk of TdP using ECG data. Second, we developed a new interpretability algorithm named Evocclusion, that enables a better understanding of the neural network’s decision process. This algorithm aims to provide human readable insights into the model’s predictions, leading to increased trust among clinicians and specialists. Third, we present two main frameworks developed to improve ECG analysis and the interpretability method. A crucial aspect of ECG analysis is signal quality. Therefore, we propose a new method using a denoising autoencoder to significantly remove noise from the ECG data and partially recover the waveform from alterations. This technique improves the reliability of the input data for subsequent analysis and ensures that the neural networks have access to high quality information. We also developed neural networks to segment the ECG and extract beats, P and T waves, and QRS complexes. These segmentation results enable a deeper understanding of the ECG components and facilitate further analysis. Additionally, we provide a method to assess a quality score vector of the ECG, enabling us to focus on parts of the signal that have a good quality score. This approach ensures that the most reliable information is used for analysis and clinicians which reduces the risk of false positives and negatives. This research seeks to enhance trust in artificial intelligence, leading to better automation of complex tasks in medicine and beyond, ultimately improving patient outcomes

Action detection is an attractive area for researchers in computer vision, healthcare, physiotherapy, psychology, and others. Intensive work has been done in this area due to its wide range of applications such as security surveillance, video tagging, Human-Computer Interaction (HCI), robotics, medical diagnosis, sports analysis, interactive gaming, and many others. After the deep learning booming results in computer vision tasks like image classification, many researchers have tried to extend the success of deep learning models to video classification and activity recognition. The research question of this thesis is to study the use of the 2D human poses extracted by a DNN-based model from RGB frames only, for the online activity detection task and comparing it with the state of the art solutions that utilize the human 3D skeletal data extracted by a depth sensor as an input. At the same time, this work showed the importance of input pre-processing and filtering on improving the performance of the online human activity detector. Detecting gym exercises and counting the repetitions in real-time using the human skeletal data versus the 2D poses have been studied in-depth in this work. The contributions of this work are as follows: 1) generating RGB-D dataset for a set of gym exercises, 2) proposing a novel real-time skeleton-based Double Representational RNN (DR-RNN) network architecture for the online action detection, 3) Demonstrating the ability of the proposed model to achieve satisfiable results using pose estimation models applied on RGB frames, 4) introducing a novel learnable exponential filter for the online low latency filtering applications.

Nanopore sequencing, a recently developed methodfor DNA sequencing, involves applying a constant electricfield over a membrane and translocating single-stranded DNAmolecules through membrane pores. This results in an electricalsignal, which is dependent on the structure of the DNA. The aimof this project is to train and evaluate a non-causal temporalconvolution neural network in order to accurately translate suchelectrical raw signal into the corresponding nucleotide sequence.The training dataset is sampled from the E. coli bacterial genomeand the phage Lambda virus. We implemented and evaluatedseveral different temporal convolutional architectures. Using anetwork with five residual blocks with five convolutional layersin each block yields maximum performance, with a predictionaccuracy of 76.1% on unseen test data. This result indicates thata temporal convolution network could be an effective way tosequence DNA data.
Nanopore sequencing är en nyligen utvecklad metod för DNA-sekvensering som innebär att man applicerar ett konstant elektriskt fält över ett membran och translokerar enkelsträngade DNA-molekyler genom membranporer. Detta resulterar i en elektrisk signal som beror på DNA-strukturen.  Målet med detta projekt är att träna och evaluera icke-kausula ”temporal convolutional networks” som ska kunna översätta denna ofiltrerade elektriska signalen till den motsvarande nukleotidsekvensen. Träningsdatan är ett urval av genomen från bakterien E. coli och viruset phage Lambda. Vi implementerade och utvärderade ett antal olika nätverksstrukturer. Ett nätverk med fem residuala block med fem faltande lager i varje block gav maximal prestation, med en precision på 76.1% på testdata. Detta resultat indikerar att ett ”temporal convolution network” skulle kunna vara ett effektivt sätt att sekvensera DNA.
Kandidatexjobb i elektroteknik 2020, KTH, Stockholm

Il presente lavoro di tesi riguarda lo studio e l'impiego di architetture neurali profonde (nello specifico stacked denoising auto-encoder) per la definizione di un modello previsionale di serie temporali. Il modello implementato è stato applicato a dati industriali riguardanti un impianto fotovoltaico reale, per effettuare una predizione della produzione di energia elettrica sulla base della serie temporale che lo caratterizza. I risultati ottenuti hanno evidenziato come la struttura neurale profonda contribuisca a migliorare le prestazioni di previsione di strumenti statistici classici come la regressione lineare multipla.

Ingången till detta arbete är att författaren upplever framtagning av material att använda i undervisningen som en mycket tidkrävande process. För lärare som börjar undervisa i en ny kurs blir detta extra tydligt när allt material ska tas fram. Med material menas planeringar, lektionsinnehåll, instuderingsuppgifter och bedömningsmaterial. Författaren har förförståelsen att lärare skulle vinna på att samarbeta och dela material mellan sig. I arbetet undersöks, genom en enkätundersökning, hur gymnasielärare i teknik gör när de tar fram material till en ny kurs eller utvecklar materialet till en kurs. Vidare undersöks vilka eventuella hinder det finns för samarbete och vilket material lärare helst vill få tillgång till från kollegor. Resultatet visar att den absoluta majoriteten av lärarna i den undersökta gruppen inte ser några hinder att dela sitt material. Gymnasielärarna i teknik delar med sig. Det kan dock finnas praktiska hinder för om att läraren är ensam på sin skola att undervisa i ämnet, finns det ingen att samarbeta med. Eller att det inte finns någon gemensam lättanvänd Community, en mötesplats på internet, att dela på. I arbetet förs också en diskussion kring möjligheter med digitalisering av material.

Artificial neural networks have been gaining attention in recent years due to theirimpressive ability to map out complex nonlinear relations within data. In this report,an attempt is made to use a Long short-term memory neural network for detectinganomalies within electrocardiographic records. The hypothesis is that if a neuralnetwork is trained on records of normal ECGs to predict future ECG sequences, it isexpected to have trouble predicting abnormalities not previously seen in the trainingdata. Three different LSTM model configurations were trained using records fromthe MIT-BIH Arrhythmia database. Afterwards the models were evaluated for theirability to predict previously unseen normal and anomalous sections. This was doneby measuring the mean squared error of each prediction and the uncertainty of over-lapping predictions. The preliminary results of this study demonstrate that recurrentneural networks with the use of LSTM units are capable of detecting anomalies.

Oggigiorno i dispositivi mobili sono l’artefatto tecnologico a maggior contatto con le persone. Pieni di sensori in grado di percepire l’ambiente circostante e dalla notevole potenza computazionale, risultano essere l’ambiente perfetto per creare applicazioni in grado di predire le azioni future e di evolvere in base alle continue scelte dell’utilizzatore. Negli ultimi anni si è fatto sempre più prorompente, nell’ambito dell’intelligenza artificiale, il Reinforcement Learning. Molte conoscenze matematiche e di programmazione sono necessarie per sfruttare al meglio questa famiglia di algoritmi all’interno delle proprie applicazioni. In questa tesi presentiamo DroidForce, un middleware per lo sfruttamento semplificato nelle proprie applicazioni di agenti di Reinforcement Learning. DroidForce espone una serie di API intuitive per la creazione e l’allenamento di questi agenti. Questo permetterà allo sviluppatore di potersi concentrare sulla logica applicativa, sfruttando la potenza del Reinforcement Learning come una black box aggiungendo poche linee di codice. Implementeremo DroidForce come una libreria Android e attraverso una applicazione demo mostreremo che rappresenta una soluzione efficiente ed efficace per integrare il Reinforcement Learning all’interno dei dispositivi mobili.

Diploma thesis is aimed to trainable image segmentation using deep neural networks. In the paper is explained the principle of digital image processing and image segmentation. In the paper is also explained the principle of artificial neural network, model of artificial neuron, training and activation of artificial neural network. In practical part of the paper is created an algorithm of sliding window to generate sub-images from image from magnetic rezonance. Generated sub-images are used to train, test and validate of the model of neural network. In practical part of the paper si created the model of the artificial neural network, which is used to trainable image segmentation. Model of the neural network is created using the Deeplearning4j library and it is optimized to parallel training using Spark library.

碩士
開南大學
資訊學院碩士在職專班
106
Biometrics and Artificial Intelligence are important infrastructures for the development of science and technology in various countries in the future. Fingerprint Identification is the most widely used and long-term application in biometric technology. Fingerprint Identification Technology has been quite mature so far, but fingerprint identification has mostly been used to explore Single Fingerprint. Recognition, the discussion of two or more Overlapped Fingerprints is relatively rare, but overlapping fingerprints is common in many criminal cases. Overlapping fingerprints have the phenomenon that the fingerprint lines interfere or cover each other. The complexity of the overlapping fingerprints is much higher than that of a single fingerprint. The difficulty in making judgments and understanding is also relatively high. At present, the recognition of overlapping fingerprints must be based on well-trained training. The personnel manually separated the fingerprint overlapping area from the non-overlapping area, which is a problem in personnel training and processing timeliness. Since the computer defeated the world chess king, artificial intelligence is the most concerned computer technology topic now, and the most important technology of artificial intelligence is Deep Learning.The main purpose of this paper is to apply the deep learning to automatically mark the overlapped and nonoverlapped regions of overlapped fingerprints.In this way, the overlapping area and the nonoverlapped area can be separated.In the case of identification of overlapping fingerprints that can be applied in criminal cases.I hope to contribute to and help with the collection of evidence in criminal cases and the speed with which cases can be solved.

碩士
元智大學
資訊工程學系
106
In advanced driver assistance system (ADAS), non-timely braking action is one of the important issues because it makes drivers exposed to a terrible and dangerous driving environment. For this reason, predicting driver braking action early and accurately must appear up, which can lower the potential of unsafe driving behavior and provide drivers more time to react. In this paper, we have sensory fusion data source from inside and outside of the car and our proposed multi-layer deep learning architecture (CBL) to predict braking action, which consists of convolutional neural network (CNN) and bidirectional long short-term memory units (BL) while CNN is good at extracting driving characteristic and BL is useful for keeping time-series data. The result points that the CBL performs much better than the other two architectures: bidirectional LSTM (BL) and uni-LSTM (UL) based on the high accuracy and f-score, and it also shows leave-one-out cross validation for drivers and many interesting differences in the speed, turning and intersection from time -5s to 0s.

Relatório de estágio apresentado à Universidade Trás-Os-Montes e Alto Douro para cumprimento dos requisitos necessários à obtenção do grau de Mestre em Engenharia Informática
O presente relatório descreve o trabalho desenvolvido no contexto do Estágio do Mestrado em Engenharia Informática da Universidade de Trás-os-Montes e Alto Douro realizado na empresa Whymob. O principal objetivo do trabalho consiste no desenvolvimento de um sistema de monitorização da desinfeção das mãos, por parte do corpo clínico em enfermarias hospitalares. O sistema proposto baseia-se em visão por computador para monitorizar o cumprimento da desinfeção hospitalar e o controlo dos 5 momentos da higienização das mãos, de maneira que seja possível combater as infeções hospitalares e, com isto, a redução do uso de antibióticos. Com este propósito, foram aplicadas técnicas de visão por computador e de deep learning, na análise de dados recolhidos por um sensor RGB-D para verificar se os funcionários hospitalares realizam a desinfeção das mãos de forma correta. A análise consiste na deteção da pose dos profissionais, deteção das mãos dos funcionários, deteção dos dispensadores de desinfetante e na identificação das categorias profissionais de cada elemento do corpo clínico. No que toca aos testes dos componentes do sistema, foi usado um cenário idêntico ao que se encontra na enfermaria hospitalar, sendo testado cada componente com um número variável de pessoas, variando esse número entre 1 e 3. Os resultados obtidos destes testes revelaram-se satisfatórios, sendo que o sistema conseguiu realizar a deteção dos elementos de interesse da enfermaria, bem como as ações realizadas por parte dos profissionais para alteração dos momentos de desinfeção.
This report describes the work carried out in the context of the master’s degree in Computer Engineering at the University of Trás-os-Montes and Alto Douro carried out at the company Whymob. The main objective of the work is the development of a system for monitoring the disinfection of hands, by the clinical staff in hospital wards. The proposed system is based on computer vision to monitor compliance with hospital disinfection and the control of the 5 moments of hand hygiene, so that it is possible to combat hospital infections and, with this, the reduction in the use of antibiotics. For this purpose, computer vision and deep learning techniques were applied in the analysis of data collected by an RGB-D sensor to verify that hospital employees perform the correct hand disinfection. The analysis consists of detecting the pose of the professionals, detecting the employees' hands, detecting the disinfectant dispensers and identifying the professional categories of each element of the clinical staff. About the testing of the system components, a scenario identical to that found in the hospital ward was used, with each component being tested with a variable number of people, varying between 1 and 3. The results obtained from these tests were revealed satisfactory, since the system was able to detect the elements of interest to the infirmary, as well as the actions taken by the professionals to change the disinfection moments.

Significant advances in fluorescence imaging techniques enable life scientists today to gain insights into biological systems at an unprecedented scale. The interpretation of image features in such bioimage datasets and their subsequent quantitative analysis is referred to as bioimage analysis. A substantial proportion of bioimage analyses is still performed manually by a human expert - a tedious process that is long known to be subjective. Particularly in tasks that require the annotation of image features with a low signal-to-noise ratio, like in fluorescence images of tissue samples, the inter-rater agreement drops. However, like any other scientific analysis, also bioimage analysis has to meet the general quality criteria of quantitative research, which are objectivity, reliability, and validity. Thus, the automation of bioimage analysis with computer-aided approaches is highly desirable. Albeit conventional hard-coded algorithms are fully unbiased, a human user has to set its respective feature extraction parameters. Thus, also these approaches can be considered subjective. Recently, deep learning (DL) has enabled impressive advances in computer vision research. The predominant difference between DL and conventional algorithms is the capability of DL models to learn the respective task on base of an annotated training dataset, instead of following user-defined rules for feature extraction. This thesis hypothesized that DL can be used to increase the objectivity, reliability, and validity of bioimage analyses, thus going beyond mere automation. However, in absence of ground truth annotations, DL models have to be trained on manual and thus subjective annotations, which could cause the model to incorporate such a bias. Moreover, model training is stochastic and even training on the same data could result in models with divergent outputs. Consequently, both the training on subjective annotations and the model-to-model variability could impair the quality of DL-based bioimage analyses. This thesis systematically assessed the impacts of these two limitations experimentally by analyzing fluorescence signals of a protein called cFOS in mouse brain sections. Since the abundance of cFOS correlates with mouse behavior, behavioral analyses could be used for cross-validation of the bioimage analysis results. Furthermore, this thesis showed that pooling the input of multiple human experts during model training and integration of multiple trained models in a model ensemble can mitigate the impact of these limitations. In summary, the present study establishes guidelines for how DL can be used to increase the general quality of bioimage analyses
Fortschritte in den Methoden der fluoreszenz-basierten Bildgebung ermöglichen Biowissenschaftlern heutzutage noch nie dagewesene Einblicke in biologische Systeme. Die Interpretation sowie die anschließende quantitative Analyse von Bildelementen in biologischen Bilddatensätzen wird in der Wissenschaft als bioimage analysis bezeichnet. Ein wesentlicher Anteil der bioimage analysis wird noch immer von Experten per Hand durchgeführt - ein mühsamer Prozess, von dem man seit langem weiß, dass er subjektiv ist. Besonders bei Aufgabestellungen, welche die Annotierung von Bildelementen mit einem geringen Signal-Rausch-Verhältnis erfordern, wie es beispielsweise bei Fluoreszenzbildern von Gewebeproben der Fall ist, sinkt die Übereinstimmung zwischen den Bewertungen mehrerer Experten. Genauso wie jede andere wissenschaftliche Analyse, muss jedoch auch die bioimage analysis den generellen Qualitätskriterien quantitativer Forschung gerecht werden. Dies sind Objektivität, Zuverlässigkeit und Validität. Die Automatisierung der bioimage analysis mit Hilfe von computer-basierten Ansätzen ist somit erstrebenswert. Konventionelle, hartkodierte Algorithmen sind zwar vollkommen unvoreingenommen, jedoch legt ein menschlicher Benutzer jene Parameter fest, die der Algorithmus für die Extraktion der relevanten Bildelemente nutzt. Aus diesem Grund sind auch diese Ansätze zumindest partiell subjektiv. In den letzten Jahren hat Deep learning (DL) zu beeindruckenden Fortschritten auf dem Forschungsgebiet der computer vision beigetragen. Der vorherrschende Unterschied zwischen DL und konventionellen Algorithmen besteht darin, dass DL Modelle in der Lage sind die jeweilige Aufgabe auf Grundlage eines annotierten Trainingsdatensatzes zu lernen, anstatt starr den Parametern zu folgen, die der Benutzer für die Extraktion der relevanten Bildelemente vorgegeben hat. In dieser Dissertation wurde die Hypothese untersucht, ob DL, neben der Möglichkeit der automatischen Bildanalyse, auch dazu genutzt werden kann die Objektivität, die Zuverlässigkeit und die Validität der Bildanalyse zu verbessern. Ohne eine objektive Referenzannotierung muss das Training der DL Modelle jedoch auf händisch erstellten und somit also subjektiven Annotierungen durchgeführt werden. Theoretisch könnte dies dazu führen, dass das DL-Modell diese Vorgeingenommenheit übernimmt. Außerdem unterliegt das Training der Modelle stochastischen Prozessen und selbst Modelle, die auf den gleichen Trainingsdaten trainiert wurden, könnten sich danach in ihren ausgegeben Analysen unterscheiden. Demzufolge könnten also sowohl das Training auf subjektiven Annotierungen als auch die Variabilität von Modell zu Modell die Qualität der DL-basierten Analyse von biologischen Bilddaten beeinträchtigen. In dieser Dissertation werden die Einflüsse von diesen beiden Limitierungen auf Grundlage von experimentellen Daten untersucht. In den experimentellen Bilddaten werden Fluoreszenzsignale des Proteins cFOS in Hirnschnitten von Mäusen dargestellt und hier repräsentativ untersucht. Da das Vorkommen von cFOS mit dem Verhalten der Mäuse korreliert, kann die Analyse des Verhaltens der Mäuse zur Kreuzvalidierung der Analyse der biologischen Bilddaten herangezogen werden. Die Daten dieser Dissertation zeigen, dass die Integration mehrerer Experten in das Training eines Modells sowie die Integration mehrerer trainierter Modelle in ein Modell-Ensemble das Risiko einer subjektiven oder nicht reproduzierbaren Bildanalyse abschwächen können. Diese Arbeit etabliert Richtlinien dafür, wie DL verwendet werden kann, um die generelle Qualität der Analyse biologischer Bilddaten zu erhöhen

The outbreak of COVID-19 is one of the major pandemics faced by the world ever and the World Health Organization (WHO) had declared it as the deadliest virus outbreak in recent times. Due to its incubation period, predicting or identifying the paints had become a tough job and thus, the impact is on a large scale. Most of the countries were affected with Coronavirus since December 2019 and the spread is still counting. Irrespective of the preventive measures being promoted on various media, still the speculations and rumors about this outbreak are peaks, that too particular with the social media platforms like Facebook and Twitter. Millions of posts or tweets are being posted on social media via various apps and due to this, the accuracy of news has become unpredictable, and further, it has increased panic among the people. To overcome these issues, a clear classification or categorization of the posts or tweets should be done to identify the accuracy of the news and this can be done by using the basic sentiment analysis technique of data sciences and machine learning. In this project, Twitter will be considered as the social media platform and the millions of tweets will be analyzed for aspect mining to categorize them into positive, negative, and neutral tweets using the NLP techniques. SVM and Naive Bayes approach of machine learning and this model will be developed.

Бублик А. О. Нейромережна система розпізнавання емоцій на обличчі людини : кваліфікаційна робота магістра спеціальності 121 «Інженерія програмного забезпечення» / наук. керівник Ю. О. Лимаренко. Запоріжжя : ЗНУ, 2020. 108 c.
UA : Мета роботи полягає у дослідженні та вивченні методів розпізнавання обличчя людини та підходів до аналізу емоцій людини,порівняння їх особ-ливостей, перевірка можливостей застосування і створення нейромережної системи для розпізнавання емоцій людини в режимі реального часу.Досліджено методи і конкуруючі сучасні системи розпізнавання емоцій з обличчя людини, їх проблематикуі можливості розробки і використання системи. Порівняно методи виділення обличчя людини і методи розпізнаван-ня емоцій людини. Спроектовано та реалізовано згорткову нейронну мережу на мові програмування Python.Створено веб-застосунок, що дозволяє розпі-знавати емоції на обличчі людини в режимі реального часу завдяки створеній нейронній мережі.
EN : The aim of the research is to study the methods of human face recognition and approaches to the analysis of human emotions, comparing their features, testing the application and creating a neural network system for recognizing human emotions in real time..Methods and competing modern systems of emotions recognition from the personfacial expressions, their problems and possibilities of development and use of system are investigated. The methods of detectinga person's face and methods of recognizing human emotions are compared. A convolutional neural network in the Python programming language was designed and implemented. Created a web application that allows userto recognize emotions in real time due tothe created neural network.

En raison de la nature cohérente du signal RADAR à synthèse d’ouverture (RSO), les images RSO polarimétriques (RSOPOL) sont affectées par le bruit de chatoiement. L’effet du chatoiement peut être sévère au point de rendre inutilisable la donnée RSOPOL. Ceci est particulièrement vrai pour les données à une vue qui souffrent d’un chatoiement très intense.Un filtrage du bruit est nécessaire pour améliorer l’estimation des paramètres polarimétriques pouvant être calculés à partir de ce type de données. Cette opération constitue une étape importante dans le traitement et l’analyse des images RSOPOL. Récemment une nouvelle approche est apparue en traitement de données visant la solution d’une multitude de problèmes dont le filtrage, la restauration d’images, la reconnaissance de la parole, la classification ou la segmentation d’images. Cette approche est l’apprentissage profond et les réseaux de neurones à convolution (RNC). Des travaux récents montrent que les RNC sont une alternative prometteuse pour le filtrages des images RSO. En effet par leur capacité d’apprendre un modèle optimal de filtrage, ils tendent à surpasser les approches classiques du filtrage sur les images RSO. L’objectif de cette présente étude est d’analyser et d’évaluer l’efficacité du filtrage par RNC sur des données RSOPOL simulées et sur des images satellitaires RSOPOL RADARSAT-2, ALOS/PalSAR et GaoFen-3 acquises sur la région urbaine de San Francisco (Californie). Des modèles inspirés de l’architecture d’un RNC utilisé notamment en Super-résolution ont été adaptés pour le filtrage de la matrice de cohérence polarimétrique. L’effet de différents paramètres structuraux de l’architecture des RNC sur le filtrage ont été analysés, parmi ceux-ci on retrouve entre autres la profondeur du réseau (le nombre de couches empilées), la largeur du réseau (le nombre de filtres par couches convolutives) et la taille des filtres de la première couche convolutive. L’apprentissage des modèles a été effectué par la rétropropagation du gradient de l’erreur en utilisant 3 ensembles de données qui simulent la polarimétrie une vue des diffuseurs selon les classes de Cloude-Pottier. Le premier ensemble ne comporte que des zones homogènes.Les deux derniers ensembles sont composés de simulations en patchwork dont l’intensité locale est simulée par des images de texture et de cibles ponctuelles ajoutées au patchwork dans le cas du dernier ensemble. Les performances des différents filtres par RNC ont été mesurées par des indicateurs comprenant l’erreur relative sur l’estimation de signatures polarimétriques et des paramètres de décomposition ainsi que des mesures de distorsion sur la récupération des détails importants et sur la conservation des cibles ponctuelles. Les résultats montrent que le filtrage par RNC des données polarimétriques est soit équivalent ou nettement supérieur aux filtres conventionnellement utilisées en polarimétrie.Les résultats des modèles les plus profonds obtiennent les meilleures performances pour tous les indicateurs sur l’ensemble des données homogènes simulées. Dans le cas des données en patchwork, les résultats pour la restauration des détails sont nettement favorables au filtrage par RNC les plus profonds.L’application du filtrage par RNC sur les images satellitaires RADARSAT-2,ALOS/PalSAR ainsi GaoFen-3 montre des résultats comparables ou supérieurs aux filtres conventionnels. Les meilleurs résultats ont été obtenus par le modèle à 5 couches cachées(si on ne compte pas la couche d’entrée et de sortie), avec 8 filtres 3×3 par couche convolutive, sauf pour la couche d’entrée où la taille des filtres étaient de 9×9. Par contre,les données d’apprentissage doivent être bien ajustées à l’étendue des statistiques des images polarimétriques réelles pour obtenir de bon résultats. Ceci est surtout vrai au niveau de la modélisation des cibles ponctuelles dont la restauration semblent plus difficiles.
Due to the coherent nature of the Synthetic Aperture Radar (SAR) signal, polarimetric SAR(POLSAR) images are affected by speckle noise. The effect of speckle can be so severe as to render the POLSAR data unusable. This is especially true for single-look data that suffer from very intense speckle. Noise filtering is necessary to improve the estimation of polarimetric parameters that can be computed from this type of data. This is an important step in the processing and analysis of POLSAR images. Recently, a new approach has emerged in data processing aimed at solving a multi-tude of problems including filtering, image restoration, speech recognition, classification orimage segmentation. This approach is deep learning and convolutional neural networks(CONVNET). Recent works show that CONVNET are a promising alternative for filtering SAR images. Indeed, by their ability to learn an optimal filtering model only from the data, they tend to outperform classical approaches to filtering on SAR images. The objective of this study is to analyze and evaluate the effectiveness of CONVNET filtering on simulated POLSAR data and on RADARSAT-2, ALOS/PalSAR and GaoFen-3 satellite images acquired over the San Francisco urban area (California). Models inspired by the architecture of a CONVNET used in particular in super-resolution have been adapted for the filtering of the polarimetric coherency matrix. The effect of different structural parameters of theCONVNET architecture on filtering were analyzed, among which are the depth of the neural network (the number of stacked layers), the width of the neural network (the number of filters per convoluted layer) and the size of the filters of the first convolution layer. The models were learned by backpropagation of the error gradient using 3 datasets that simulate single-look polarimetry of the scatterers according to Cloude-Pottier classes. The first dataset contains only homogeneous areas. The last two datasets consist of patchwork simulations where local intensity is simulated by texture images and point target are added to the patchwork in the case of the last dataset. The performance of the different filters by CONVNET was measured by indicators including relative error on the estimation of polarimetric signatures and decomposition parameters as well as distortion measurements on the recovery of major details and on the conservation of point targets.The results show that CONVNET filtering of polarimetric data is either equivalent or significantly superior to conventional polarimetric filters. The results of the deepest models obtain the best performance for all indicators over the simulated homogeneous dataset. Inthe case of patchwork dataset, the results for detail restoration are clearly favourable to the deepest CONVNET filtering. The application of CONVNET filtering on RADARSAT-2, ALOS/PalSAR andGaoFen-3 satellite images shows results comparable or superior to conventional filters. The best results were obtained by the 5 hidden layers model (not counting the input and outputlayers), with 8 filters 3×3 per convolutional layer, except for the input layer where the filtersize was 9×9. On the other hand, the training data must be well adjusted to the statistical range of the real polarimetric images to obtain good results. This is especially true when modeling point targets that appear to be more difficult to restore.