This report describes an approach of creating a system identifying fruit and vegetables in the retail market using images captured with a video camera at- tached to the system. The system helps the customers to label desired fruits and vegetables with a price according to its weight. The purpose of the sys- tem is to minimize the number of human computer interactions, speed up the identification process and improve the usability of the graphical user interface compared to existing systems. To accomplish creating a system improving these properties, an idea of implementing machine learning to identify the products aroused. Instead of assigning the responsibility to the user, who usually iden- tify the products manually, the responsibility is given to a computer. To classify an object, different convolutional neural networks have been tested and retrained. The networks have been retrained on data sets collected from ImageNet. To improve the accuracy, the networks have also been retrained on images where the background environment is similar to the environment the networks are supposed to perform in. The networks tested in this report are MobileNet and Inception. The networks have different propagation time and varies in accuracy. MobileNet performs the classification about seven times faster than Inception, but Inception gives more accurate results. To improve the systems further, usability testing has been performed on the graphical user interface of existing system and resulted system. To test the usability, a heuristic evaluation has been performed in combination of a second test produced by the authors. The tests concluded that the resulted system was more user friendly compared to existing systems. The hardware of the system constitutes of a Raspberry Pi, camera, display, load cell and a case. The software includes Python-code to label an image, a graphical user interface to interact with the user and a server created with Node.js. The graphical user interface has been programmed with JavaScript supplemented with the React library. To conclude, implementing convolutional neural networks to classify images and developing a new user interface resulted in a faster identification process together with fewer usability flaws.

Using Convolutional Neural Networks for classification of images and for localization and detection of objects in images is becoming increasingly popular. Within radiology a huge amount of image data is produced and meta data containing information of what the images depict is currently added manually by a radiologist. To aid in streamlining physician’s workflow this study has investigated the possibility to use Convolutional Neural Networks (CNNs) that are pre-trained on natural images to automatically detect the presence and location of multiple organs and body-parts in medical CT images. The results show promise for multiclass classification with an average precision 89.41% and average recall 86.40%. This also confirms that a CNN that is pre-trained on natural images can be succesfully transferred to solve a different task. It was also found that adding additional data to the dataset does not necessarily result in increased precision and recall or decreased error rate. It is rather the type of data and used preprocessing techniques that matter.

本篇文章使用了增強學習與捲積深度學習結合的DQCN模型製作交易系統，希望藉由此交易系統能自行判斷是否買賣ETF，由於ETF屬於穩定性高且手續費高的衍生性金融商品，所以該系統不即時性的做買賣，採用每二十個開盤日進行一次買賣，並由這20個開盤日進行買賣的預測，希望該系統能最大化我們未來的報酬。 DQN是一種增強學習的模型，並在其中使用深度學習進行動作價值的預測，利用增強學習的自我更新動作價值的機制，再用深度學習強大的學習能力成就了人工智慧，並在其取得良好的成效。 / In this paper, we used DCQN model, which is combined with reinforcement learning and CNN to train a trading system and hope the trading system could judge whether buy or sell ETFs. Since ETFs is a derivative financial good with high stability and related fee, the system does not perform real-time trading and it performs every 20 trading day. The system predicts value of action based on data in the last 20 opening days to maximize our future rewards. DQN is a reinforcement learning model, using deep learning to predict value of actions in model. Combined with the RL's mechanism, which updates value of actions, and deep learning, which has a strong ability of learning, to finish an artificial intelligence. We got a perfect effect.

Now and then train accidents occur. Collisions between trains and objects such as animals, humans, cars, and fallen trees can result in casualties, severe damage on the train, and delays in the train traffic. Thus, train collisions are a considerable problem with consequences affecting society substantially. The company Termisk Systemteknik AB has on commission by Rindi Solutions AB investigated the possibility to detect anomalies on the railway using a trainmounted thermal imaging camera. Rails are also detected in order to determine if an anomaly is on the rail or not. However, the rail detection method does not work satisfactory at long range. The purpose of this master’s thesis is to improve the previous rail detector at long range by using machine learning, and in particular deep learning and a convolutional neural network. Of interest is also to investigate if there are any advantages using cross-modal transfer learning. A labelled dataset for training and testing was produced manually. Also, a loss function tailored to the particular problem at hand was constructed. The loss function was used both for improving the system during training and evaluate the system’s performance during testing. Finally, eight different approaches were evaluated, each one resulting in a different rail detector. Several of the rail detectors, and in particular all the rail detectors using crossmodal transfer learning, perform better than the previous rail detector. Thus, the new rail detectors show great potential to the rail detection problem.

Are foreign policy officials responsive to policy preferences of the mass media and the public in making their decisions? That question has dogged scholars for decades but there has been little agreement among them on what is the true nature of mass media- and public opinion-foreign policy link. In terms of mass media impact, there are two media theories which dominate the debate. First, the CNN Effect theory claims that, by their nature, the mass media have the power to compel policy officials to adopt their policy preferences. Second, the Manufacturing Consent theory counters with the claim that foreign policy is too serious a matter for officials to yield to mass media demands. Scholars are similarly divided on the impact of public opinion on foreign policy. Lacking in almost all the known studies is an appreciation that foreign policy emerges out of a process involving policy stages. These policy stages have different characteristics. In addition to the nature of those stages in themselves, relationships between policy actors - including the mass media, the public and officials - are different in those stages. Officials tend to react differently at each stage of policy when pressured by the mass media and public opinion. Therefore, in this study, I propose that we will have a better understanding of mass media and public opinion influence on foreign policy officials if we study official responsiveness or sensitivity at the stages of the foreign policy process - policy initiation, policy implementation and policy review. I further argue that official responsiveness to mass media and public opinion depends largely on the stage of policy. For this research, I carried out a case study of Britain's war with Iraq in 2003 to test my theory. Principally, I tried to answer the question: Does foreign policy officials' responsiveness to mass media and public opinion depend on the stage of policy? I found that official response to the mass media and public opinion was not as precise as suggested by the dominant camps in the debate. More importantly, Official response to mass media and public opinion varied in the stages of policy. Specifically, I found that British officials were most responsive to mass media and public opinion at the policy initiation stage, very unresponsive at the implementation stage and even more unresponsive at the policy review stage. As a result of the variations in official responsiveness at the stages, I argue that there is a need to re-evaluate the way we study mass media- and public opinion-foreign policy link. To better understand the impact of the mass media and public opinion on foreign policy, I conclude that we need to examine how policy actors interact at different stages of the foreign policy process.

Notre travail est présenté en trois parties indépendantes. Tout d'abord, nous proposons trois heuristiques d'apprentissage actif pour les réseaux de neurones profonds : Nous mettons à l'échelle le `query by committee' , qui agrège la décision de sélectionner ou non une donnée par le vote d'un comité. Pour se faire nous formons le comité à l'aide de différents masques de dropout. Un autre travail se base sur la distance des exemples à la marge. Nous proposons d'utiliser les exemples adversaires comme une approximation de la dite distance. Nous démontrons également des bornes de convergence de notre méthode dans le cas de réseaux linéaires. L’usage des exemples adversaires ouvrent des perspectives de transférabilité d’apprentissage actif d’une architecture à une autre. Puis, nous avons formulé une heuristique d'apprentissage actif qui s'adapte tant au CNNs qu'aux RNNs. Notre méthode sélectionne les données qui minimisent l'énergie libre variationnelle. Dans un second temps, nous nous sommes concentrés sur la distance de Wasserstein. Nous projetons les distributions dans un espace où la distance euclidienne mimique la distance de Wasserstein. Pour se faire nous utilisons une architecture siamoise. Également, nous démontrons les propriétés sous-modulaires des prototypes de Wasserstein et comment les appliquer à l'apprentissage actif. Enfin, nous proposons de nouveaux outils de visualisation pour expliquer les prédictions d'un CNN sur du langage naturel. Premièrement, nous détournons une stratégie d'apprentissage actif pour confronter la pertinence des phrases sélectionnées aux techniques de phraséologie les plus récentes. Deuxièmement, nous profitons des algorithmes de déconvolution des CNNs afin de présenter une nouvelle perspective sur l'analyse d'un texte. / Our work is presented in three separate parts which can be read independently. Firstly we propose three active learning heuristics that scale to deep neural networks: We scale query by committee, an ensemble active learning methods. We speed up the computation time by sampling a committee of deep networks by applying dropout on the trained model. Another direction was margin-based active learning. We propose to use an adversarial perturbation to measure the distance to the margin. We also establish theoretical bounds on the convergence of our Adversarial Active Learning strategy for linear classifiers. Some inherent properties of adversarial examples opens up promising opportunity to transfer active learning data from one network to another. We also derive an active learning heuristic that scales to both CNN and RNN by selecting the unlabeled data that minimize the variational free energy. Secondly, we focus our work on how to fasten the computation of Wasserstein distances. We propose to approximate Wasserstein distances using a Siamese architecture. From another point of view, we demonstrate the submodular properties of Wasserstein medoids and how to apply it in active learning. Eventually, we provide new visualization tools for explaining the predictions of CNN on a text. First, we hijack an active learning strategy to confront the relevance of the sentences selected with active learning to state-of-the-art phraseology techniques. These works help to understand the hierarchy of the linguistic knowledge acquired during the training of CNNs on NLP tasks. Secondly, we take advantage of deconvolution networks for image analysis to present a new perspective on text analysis to the linguistic community that we call Text Deconvolution Saliency.

<div>This thesis consists of four sections which are related with four research projects.</div><div><br></div><div>The first section is about Page Classification. In this section, we extend our previous approach which could classify 3 classes of pages: Text, Picture and Mixed, to 5 classes which are: Text, Picture, Mixed, Receipt and Highlight. We first design new features to define those two new classes and then use DAG-SVM to classify those 5 classes of images. Based on the results, our algorithm performs well and is able to classify 5 types of pages.</div><div><br></div><div>The second section is about Fading Detection. In this section, we develop an algorithm that can automatically detect fading for both text and non-text region. For text region, we first do global alignment and then perform local alignment. After that, we create a 3D color node system, assign each connected component to a color node and get the color difference between raster page connected component and scanned page connected. For non-text region, after global alignment, we divide the page into "super pixels" and get the color difference between raster super pixels and testing super pixels. Compared with the traditional method that uses a diagnostic page, our method is more efficient and effective.</div><div><br></div><div>The third section is about CNN Based Emotion Recognition. In this section, we build our own emotion recognition classification and regression system from scratch. It includes data set collection, data preprocessing, model training and testing. We extend the model to real-time video application and it performs accurately and smoothly. We also try another approach of solving the emotion recognition problem using Facial Action Unit detection. By extracting Facial Land Mark features and adopting SVM training framework, the Facial Action Unit approach achieves comparable accuracy to the CNN based approach.</div><div><br></div><div>The forth section is about Saliency Based Image Quality Assessment and Cropping. In this section, we propose a method of doing image quality assessment and recomposition with the help of image saliency information. Saliency is the remarkable region of an image that attracts people's attention easily and naturally. By showing everyday examples as well as our experimental results, we demonstrate the fact that, utilizing the saliency information will be beneficial for both tasks.</div>

Ultrasonic sensors are commonly used in automobiles to assist driving maneuvers, e.g., parking, because of their cost-effectiveness and robustness. This thesis investigated the feasibility of using an Ultrasonic Sensor Array to locate the 3D position of an object and also using the measurements from the sensor array to train a Convolutional Neural Network (CNN) to classify the objects. A simulated Ultrasonic Sensor array was built in COMSOL Multiphysics. The simulation of ultrasound used Ray Tracing technology to track the path of ultrasound rays. The readouts from the sensor array are used to formulate an optimization problem to address the 3D positioning of the object. We investigated the performance of two optimization methods in terms of the accuracy of the prediction and the efficiency of solving the problem. The average mean absolute error (MAE) and average mean squared error (MSE) of the Nelder-Mead method (without constraints) are 2.66 mm and 12.79 mm2 respectively, the average running time to predict one 3D position is 97.62 ms. The average MAE and average MSE of Powell’s method (with constraints) are 2.84 mm and 23.66 mm2 respectively, average running time to predict one 3D position is 84.68 ms. The result of Powell’s method (without constraints) is much worse than the above two, its average MAE and MSE are 24.93 mm and 7559.46 mm2, average running time is 238.30 ms. The readouts from the sensor array are also used to build eight different datasets of which the data structures are different combinations of the information from the readouts. Each of these eight data sets is used to train a CNN, and the classification accuracy of each CNN indicates that how well the data structure represents the objects. The results showed that the CNN trained by stacked time array 5×5×3 had the best classification accuracy among eight datasets, the classification accuracy on the test set is 85.05%.

In this thesis is handled the problem of mushroom detection and recognition in natural environment. Convolutional neural networks are used. The beginning of this thesis is dedicated to the theory of neural networks. Further is solved the problem of object detection and classification. Using neural network trained for classification is solved also the task of localization. Results of trained CNNs are analised.

The goal of this work is to bring automatic defect detection to the manufacturing process of plastic cards. A card is considered defective when it is contaminated with a dust particle or a hair. The main challenges I am facing to accomplish this task are a very few training data samples (214 images), small area of target defects in context of an entire card (average defect area is 0.0068 \% of the card) and also very complex background the detection task is performed on. In order to accomplish the task, I decided to use Mask R-CNN detection algorithm combined with augmentation techniques such as synthetic dataset generation. I trained the model on the synthetic dataset consisting of 20 000 images. This way I was able to create a model performing 0.83 AP at 0.1 IoU on the original data test set.