During recent years, great progress has been made within the field of deep learning, and more specifically, within neural networks. Deep convolutional neural networks (CNN) have been especially successful within image processing in tasks such as image classification and object detection. Car manufacturers, amongst other actors, are starting to realize the potential of deep learning and have begun applying it to autonomous driving. This is not a simple task, and many challenges still lie ahead. A sub-problem, that needs to be solved, is a way of automatically determining the road conditions, including the friction. Since many modern cars are equipped with cameras these days, it is only natural to approach this problem with CNNs. This is what has been done in this thesis. First, a data set is gathered which consists of 37,000 labeled road images that are taken through the front window of a car. Second, CNNs are trained on this data set to classify the friction of a given road. Gathering road images and labeling them with the correct friction is a time consuming and difficult process, and requires human supervision. For this reason, experiments are made on a second data set, which consist of 54,000 simulated images. These images are captured from the racing game World Rally Championship 7 and are used in addition to the real images, to investigate what can be gained from this. Experiments conducted during this thesis show that CNNs are a good approach for the problem of estimating the road friction. The limiting factor, however, is the data set. Not only does the data set need to be much bigger, but it also has to include a much wider variety of driving conditions. Friction is a complex property and depends on many variables, and CNNs are only effective on the type of data that they have been trained on. For these reasons, new data has to be gather by actively seeking different driving conditions in order for this approach to be deployable in practice. / Under de senaste åren har det gjorts stora framsteg inom maskininlärning, särskilt gällande neurala nätverk. Djupa neurala närverk med faltningslager, eller faltningsnätverk (eng. convolutional neural network) har framför allt varit framgångsrika inom bildbehandling i problem så som bildklassificering och objektdetektering. Biltillverkare, bland andra aktörer, har nu börjat att inse potentialen av maskininlärning och påbörjat dess tillämpning inom autonom körning. Detta är ingen enkel uppgift och många utmaningar finns fortfarande framöver. Ett delproblem som måste lösas är ett sätt att automatiskt avgöra väglaget, där friktionen ingår. Eftersom många nya bilar är utrustade med kameror är det naturligt att försöka tackla detta problem med faltningsnätverk, vilket är varför detta har gjorts under detta examensarbete. Först samlar vi in en datamängd beståendes av 37 000 bilder tagna på vägar genom framrutan av en bil. Dessa bilder kategoriseras efter friktionen på vägen. Sedan tränar vi faltningsnätverk på denna datamängd för att klassificera friktionen. Att samla in vägbilder och att kategorisera dessa är en tidskrävande och svår process och kräver mänsklig övervakning. Av denna anledning utförs experiment på en andra datamängd beståendes av 54 000 simulerade bilder. Dessa har blivit insamlade genom spelet World Rally Championship 7 där syftet är att undersöka om prestandan på nätverken kan ökas genom simulerat data och därmed minska kravet på storleken av den riktiga datamängden. De experiment som har utförts under examensarbetet visar på att faltningsnätverk är ett bra tillvägagångssätt för att skatta vägfriktionen. Den begränsande faktorn i det här fallet är datamängden. Datamängden behöver inte bara vara större, men den måste framför allt täcka in ett bredare urval av väglag och väderförhållanden. Friktion är en komplex egenskap och beror på många variabler, och faltningsnätverk är endast effektiva på den typen av data som de har tränats på. Av dessa anledningar behöver ny data samlas in genom att aktivt söka efter nya körförhållanden om detta tillvägagångssätt ska vara tillämpbart i praktiken.

Digital photography became widespread with the global use of smartphones. However, most of the captured images do not fully use the camera capabilities by storing the captured photos in a format with limited dynamic range. The subject of dynamic range expansion and reconstruction has been researched since early 2000s and recently gave rise to several new reconstruction methods using convolutional neural networks (CNNs), whose performance has not yet been comprehensively compared. By implementing and using our dynamic range reconstruction evaluation framework we compare the reconstruction quality of individual CNN-based approaches. We also implement a mobile HDR camera application and evaluate the feasibility of running the best-performing reconstruction method directly on a mobile device.

In recent years, development of Convolutional Neural Networks has enabled high performing semantic segmentation models. Generally, these deep learning based segmentation methods require a large amount of annotated data. Acquiring such annotated data for semantic segmentation is a tedious and expensive task. Within machine learning, active learning involves in the selection of new data in order to limit the usage of annotated data. In active learning, the model is trained for several iterations and additional samples are selected that the model is uncertain of. The model is then retrained on additional samples and the process is repeated again. In this thesis, an active learning framework has been applied to road segmentation which is semantic segmentation of objects related to road scenes. The uncertainty in the samples is estimated with Monte Carlo dropout. In Monte Carlo dropout, several dropout masks are applied to the model and the variance is captured, working as an estimate of the model’s uncertainty. Other metrics to rank the uncertainty evaluated in this work are: a baseline method that selects samples randomly, the entropy in the default predictions and three additional variations/extensions of Monte Carlo dropout. Both the active learning framework and uncertainty estimation are implemented in the thesis. Monte Carlo dropout performs slightly better than the baseline in 3 out of 4 metrics. Entropy outperforms all other implemented methods in all metrics. The three additional methods do not perform better than Monte Carlo dropout. An analysis of what kind of uncertainty Monte Carlo dropout capture is performed together with a comparison of the samples selected by baseline and Monte Carlo dropout. Future development and possible improvements are also discussed.

One of the biggest challenges in speech recognition today is its use on a daily basis, in which distortion and noise in the environment are present and hinder this task. In the last thirty years, hundreds of methods for noise-robust recognition were proposed, each with its own advantages and disadvantages. In this thesis, the use of Convolutional Neural Networks (CNN) as acoustic models in automatic speech recognition systems (ASR) is proposed as an alternative to the classical recognition methods based on Hidden Markov Models (HMM) without any noise-robust method applied. Experiments were performed with a audio set modified by additive and natural noises, and showed that the presented method reduces the Equal Error Rate (EER) and improves the acuracy of speech recognition in noisy environments when compared to traditional models of classifiation, indicating the robustness of the approach. / Um dos maiores desafios no reconhecimento de fala atualmente é usá-lo no contexto diário, no qual distorções no sinal da fala e ruídos no ambiente estão presentes e re- duzem a qualidade do reconhecimento. Nos últimos trinta anos, centenas de métodos para reconhecimento robusto ao ruído foram propostos, cada um com suas vantagens e desvantagens. Este trabalho propõe o uso de uma rede neural convolucional no papel de modelo acústico em sistemas de reconhecimento automático de fala,como uma alter- nativa ao métodos clássicos de reconhecimento baseado em modelos ocultos de Markov (HMM, do inglês, Hidden Markov Models) sem a aplicação de um método robusto ao ruído. Experimentos foram realizados com áudios modi ficados com ruídos aditivos e reais, e mostraram que o método proposto reduz o Equal Error Rate (EER) e aumenta a acurácia da classificação de comando de voz quando comparado a modelos tradicionais de classificação, evidenciando a robustez da abordagem apresentada.

The use of machine learning and specifically neural networks is a growing trend in software development, and has grown immensely in the last couple of years in the light of an increasing need to handle big data and large information flows. Machine learning has a broad area of application, such as human-computer interaction, predicting stock prices, real-time translation, and self driving vehicles. Large companies such as Microsoft and Google have already implemented machine learning in some of their commercial products such as their search engines, and their intelligent personal assistants Cortana and Google Assistant. The main goal of this project was to evaluate the two deep learning frameworks Google TensorFlow and Microsoft CNTK, primarily based on their performance in the training time of neural networks. We chose to use the third-party API Keras instead of TensorFlow's own API when working with TensorFlow. CNTK was found to perform better in regards of training time compared to TensorFlow with Keras as frontend. Even though CNTK performed better on the benchmarking tests, we found Keras with TensorFlow as backend to be much easier and more intuitive to work with. In addition, CNTKs underlying implementation of the machine learning algorithms and functions differ from that of the literature and of other frameworks. Therefore, if we had to choose a framework to continue working in, we would choose Keras with TensorFlow as backend, even though the performance is less compared to CNTK.

Abstract : Malignant brain tumors are the second leading cause of cancer related deaths in children under 20. There are nearly 700,000 people in the U.S. living with a brain tumor and 17,000 people are likely to loose their lives due to primary malignant and central nervous system brain tumor every year. To identify whether a patient is diagnosed with brain tumor in a non-invasive way, an MRI scan of the brain is acquired followed by a manual examination of the scan by an expert who looks for lesions (i.e. cluster of cells which deviate from healthy tissue). For treatment purposes, the tumor and its sub-regions are outlined in a procedure known as brain tumor segmentation . Although brain tumor segmentation is primarily done manually, it is very time consuming and the segmentation is subject to variations both between observers and within the same observer. To address these issues, a number of automatic and semi-automatic methods have been proposed over the years to help physicians in the decision making process. Methods based on machine learning have been subjects of great interest in brain tumor segmentation. With the advent of deep learning methods and their success in many computer vision applications such as image classification, these methods have also started to gain popularity in medical image analysis. In this thesis, we explore different machine learning and deep learning methods applied to brain tumor segmentation. / Résumé: Les tumeurs malignes au cerveau sont la deuxième cause principale de décès chez les enfants de moins de 20 ans. Il y a près de 700 000 personnes aux États-Unis vivant avec une tumeur au cerveau, et 17 000 personnes sont chaque année à risque de perdre leur vie suite à une tumeur maligne primaire dans le système nerveu central. Pour identifier de façon non-invasive si un patient est atteint d'une tumeur au cerveau, une image IRM du cerveau est acquise et analysée à la main par un expert pour trouver des lésions (c.-à-d. un groupement de cellules qui diffère du tissu sain). Une tumeur et ses régions doivent être détectées à l'aide d'une segmentation pour aider son traitement. La segmentation de tumeur cérébrale et principalement faite à la main, c'est une procédure qui demande beaucoup de temps et les variations intra et inter expert pour un même cas varient beaucoup. Pour répondre à ces problèmes, il existe beaucoup de méthodes automatique et semi-automatique qui ont été proposés ces dernières années pour aider les praticiens à prendre des décisions. Les méthodes basées sur l'apprentissage automatique ont suscité un fort intérêt dans le domaine de la segmentation des tumeurs cérébrales. L'avènement des méthodes de Deep Learning et leurs succès dans maintes applications tels que la classification d'images a contribué à mettre de l'avant le Deep Learning dans l'analyse d'images médicales. Dans cette thèse, nous explorons diverses méthodes d'apprentissage automatique et de Deep Learning appliquées à la segmentation des tumeurs cérébrales.

Tout d'abord, nous présentons des méthodes permettant d'inférer un graphe à partir de signaux, afin de modéliser le support des données à classifier. Ensuite, des translations préservant les voisinages des sommets sont identifiées sur le graphe inféré. Enfin, ces translations sont utilisées pour déplacer un noyau convolutif sur le graphe, afin dedéfinir un réseau de neurones convolutif adapté aux données d'entrée.Nous avons illustré notre méthodologie sur une base de données d'images. Sans utiliser de connaissances sur les signaux, nous avons pu inférer un graphe proche d'une grille. Les translations sur ce graphe sont proches des translations Euclidiennes, ce qui nous a permis de définir un réseau de neurones convolutif très similaire à ce que l'on aurait pu obtenir en utilisant l'information que les signaux sont des images. Ce réseau, entraîné sur les données initiales, a dépassé lesperformances des méthodes de l'état de l'art de plus de 13 points, tout en étant simple et facilement améliorable.La méthode que nous avons introduite est une généralisation des réseaux de neurones convolutifs, car ceux-ci sont des cas particuliers de notre approche quand le graphe est une grille. Nos travaux ouvrent donc de nombreuses perspectives, car ils fournissent une méthode efficace pour construire des réseaux adaptés aux données. / This manuscript sums up our work on extending convolutional neuralnetworks to irregular domains through graph inference. It consists of three main chapters, each giving the details of a part of a methodology allowing the definition of such networks to process signals evolving on graphs with unknown structures.First, graph inference from data is explored, in order to provide a graph modeling the support of the signals to classify. Second, translation operators that preserve neighborhood properties of the vertices are identified on the inferred graph. Third, these translations are used to shift a convolutional kernel on the graph in order to define a convolutional neural network that is adapted to the input data.We have illustrated our methodology on a dataset of images. While not using any particular knowledge on the signals, we have been able to infer a graph that is close to a grid. Translations on this graph resemble Euclidean translations. Therefore, this has allowed us to define an adapted convolutional neural network that is very close what one would obtain when using the information that signals are images. This network, trained on the initial data, has out performed state of the art methods by more than 13 points, while using a very simple and easily improvable architecture.The method we have introduced is a generalization of convolutional neural networks. As a matter of fact, they can be seen as aparticularization of our approach in the case where the graph is a grid. Our work thus opens the way to numerous perspectives, as it provides an efficient way to build networks that are adapted to the data.

Deep learning has been rapidly growing in recent years obtaining excellent results for many computer vision applications, such as image classification and object detection. One aspect for the increased popularity of deep learning is that it mitigates the need for hand-crafted features. This thesis work investigates deep learning as a methodology to solve the problem of autonomous collision avoidance for a small robotic car. To accomplish this, transfer learning is used with the VGG16 deep network pre-trained on ImageNet dataset. A dataset has been collected and then used to fine-tune and validate the network offline. The deep network has been used with the robotic car in a real-time manner. The robotic car sends images to an external computer, which is used for running the network. The predictions from the network is sent back to the robotic car which takes actions based on those predictions. The results show that deep learning has great potential in solving the collision avoidance problem.

Understanding biology paves the way for discovering drugs targeting deadly diseases like cancer, and microscopy imaging is one of the most informative ways to study biology. However, analysis of large numbers of samples is often required to draw statistically verifiable conclusions. Automated approaches for analysis of microscopy image data makes it possible to handle large data sets, and at the same time reduce the risk of bias. Quantitative microscopy refers to computational methods for extracting measurements from microscopy images, enabling detection and comparison of subtle changes in morphology or behavior induced by varying experimental conditions. This thesis covers computational methods for segmentation and classification of biological samples imaged by microscopy. Recent increase in computational power has enabled the development of deep neural networks (DNNs) that perform well in solving real world problems. This thesis compares classical image analysis algorithms for segmentation of bacteria cells and introduces a novel method that combines classical image analysis and DNNs for improved cell segmentation and detection of rare phenotypes. This thesis also demonstrates a novel DNN for segmentation of clusters of cells (spheroid), with varying sizes, shapes and textures imaged by phase contrast microscopy. DNNs typically require large amounts of training data. This problem is addressed by proposing an automated approach for creating ground truths by utilizing multiple imaging modalities and classical image analysis. The resulting DNNs are applied to segment unstained cells from bright field microscopy images. In DNNs, it is often difficult to understand what image features have the largest influence on the final classification results. This is addressed in an experiment where DNNs are applied to classify zebrafish embryos based on phenotypic changes induced by drug treatment. The response of the trained DNN is tested by ablation studies, which revealed that the networks do not necessarily learn the features most obvious at visual examination. Finally, DNNs are explored for classification of cervical and oral cell samples collected for cancer screening. Initial results show that the DNNs can respond to very subtle malignancy associated changes. All the presented methods are developed using open-source tools and validated on real microscopy images.

This study presents an exploration of several machine learning and image processing theories, as well as a literature review of several previous works on concrete crack detection systems. Through the literature review a system is selected and implemented with the Öresund bridge photograph collection. The selected system is a Convolutional Neural Network (CNN) using cropped (256x256x) images for input. The CNN has a total of 13 layers that were implemented as described in the paper. All parts of the implementation such as cropping, code, and testing are described in detail. This study found a final accuracy rate of 77% for the trained net. This is combined with a sliding window technique for handling larger images. An exploration of reasons for this accuracy rate is done at the end of the paper.