With the advancement in deep learning in the past few years, we are able to create complex machine learning models for detecting objects in images, regardless of the characteristics of the objects to be detected. This development has enabled engineers to replace existing heuristics-based systems in favour of machine learning models with superior performance. In this report, we evaluate the viability of using deep learning models for object detection in real-time video feeds on mobile devices in terms of object detection performance and inference delay as either an end-to-end system or feature extractor for existing algorithms. Our results show a significant increase in object detection performance in comparison to existing algorithms with the use of transfer learning on neural networks adapted for mobile use. / Utvecklingen inom djuplärning de senaste åren innebär att vi är kapabla att skapa mer komplexa maskininlärningsmodeller för att identifiera objekt i bilder, oavsett objektens attribut eller karaktär. Denna utveckling har möjliggjort forskare att ersätta existerande heuristikbaserade algoritmer med maskininlärningsmodeller med överlägsen prestanda. Den här rapporten syftar till att utvärdera användandet av djuplärningsmodeller för exekvering av objektigenkänning i video på mobila enheter med avseende på prestanda och exekveringstid. Våra resultat visar på en signifikant ökning i prestanda relativt befintliga heuristikbaserade algoritmer vid användning  av djuplärning och överförningsinlärning i artificiella neurala nätverk.

<p>Wearable devices and their ubiquitous use and deployment across multiple areas of health provide key insights in patient and individual status via big data through sensor capture at key parts of the individual’s body. While small and low cost, their limitations rest in their computational and battery capacity. One key use of wearables has been in individual activity capture. For accelerometer and gyroscope data, oscillatory patterns exist between daily activities that users may perform. By leveraging spatial and temporal learning via CNN and LSTM layers to capture both the intra and inter-oscillatory patterns that appear during these activities, we deployed data sparsification via autoencoders to extract the key topological properties from the data and transmit via BLE that compressed data to a central device for later decoding and analysis. Several autoencoder designs were developed to determine the principles of system design that compared encoding overhead on the sensor device with signal reconstruction accuracy. By leveraging asymmetric autoencoder design, we were able to offshore much of the computational and power cost of signal reconstruction from the wearable to the central devices, while still providing robust reconstruction accuracy at several compression efficiencies. Via our high-precision Bluetooth voltmeter, the integrated sparsified data transmission configuration was tested for all quantization and compression efficiencies, generating lower power consumption to the setup without data sparsification for all autoencoder configurations. </p> <p><br></p> <p>Human activity recognition (HAR) is a key facet of lifestyle and health monitoring. Effective HAR classification mechanisms and tools can provide healthcare professionals, patients, and individuals key insights into activity levels and behaviors without the intrusive use of human or camera observation. We leverage both spatial and temporal learning mechanisms via CNN and LSTM integrated architectures to derive an optimal classification architecture that provides robust classification performance for raw activity inputs and determine that a LSTMCNN utilizing a stacked-bidirectional LSTM layer provides superior classification performance to the CNNLSTM (also utilizing a stacked-bidirectional LSTM) at all input widths. All inertial data classification frameworks are based off sensor data drawn from wearable devices placed at key sections of the body. With the limitation of wearable devices being a lack of computational and battery power, data compression techniques to limit the quantity of transmitted data and reduce the on-board power consumption have been employed. While this compression methodology has been shown to reduce overall device power consumption, this comes at a cost of more-or-less information loss in the reconstructed signals. By employing an asymmetric autoencoder design and training the LSTMCNN classifier with the reconstructed inputs, we minimized the classification performance degradation due to the wearable signal reconstruction error The classifier is further trained on the autoencoder for several input widths and with quantized and unquantized models. The performance for the classifier trained on reconstructed data ranged between 93.0\% and 86.5\% accuracy dependent on input width and autoencoder quantization, showing promising potential of deep learning with wearable sparsification. </p>

Background:The background of this research lies in detecting the images from satellites. The recognition of images from satellites has become increasingly importantdue to the vast amount of data that can be obtained from satellites. This thesisaims to develop a method for the recognition of images from satellites using machinelearning techniques. Objective:The main objective of this thesis is a unique approach to recognizingthe data with a CRNN algorithm that involves image recognition in satellite imagesusing machine learning, specifically the CRNN (Convolutional Recurrent Neural Network) architecture. The main task is classifying the images accurately, and this isachieved by utilizing object classification algorithms. The CRNN architecture ischosen because it can effectively extract features from satellite images using Convolutional Blocks and leverage the great memory power of the Long Short-TermMemory (LSTM) networks to connect the extracted features efficiently. The connected features improve the accuracy of our model significantly. Method:The proposed method involves doing a literature review to find currentimage recognition models and then experimentation by training a CRNN, CNN andRNN and then comparing their performance using metrics mentioned in the thesis work. Results:The performance of the proposed method is evaluated using various metrics, including precision, recall, F1 score and inference speed, on a large dataset oflabeled images. The results indicate that high accuracy is achieved in detecting andclassifying objects in satellite images through our approach. The potential utilization of our proposed method can span various applications such as environmentalmonitoring, urban planning, and disaster management. Conclusion:The classification on the satellite images is performed using the 2 datasetsfor ships and cars. The proposed architectures are CRNN, CNN, and RNN. These3 models are compared in order to find the best performing algorithm. The resultsindicate that CRNN has the best accuracy and precision and F1 score and inferencespeed, indicating a strong performance by the CRNN. Keywords: Comparison of CRNN, CNN, and RNN, Image recognition, MachineLearning, Algorithms,You Only Look Once. Version3, Satellite images, Aerial Images, Deep Learning

Healthcare systems are constantly challenged with bottlenecks due to human-reliant operations, such as analyzing medical images. High precision and repeatability is necessary when performing a diagnostics on patients with tumors. Throughout the years an increasing number of advancements have been made using various machine learning algorithms for the detection of tumors helping to fast track diagnosis and treatment decisions. “Black Box” systems such as the complex deep learning networks discussed in this paper rely heavily on hyperparameter optimization in order to obtain the most ideal performance. This requires a significant time investment in the tuning of such networks to acquire cutting-edge results. The approach of this paper relies on implementing a state of the art deep learning framework, the nn-UNet, in order to label computed tomography (CT) images from patients with kidney cancer through semantic segmentation by feeding raw CT images through a deep architecture and obtaining pixel-wise mask classifications. Taking advantage of nn-UNet’s framework versatility, various configurations of the architecture are explored and applied, benchmarking and assorting resulting performance, including variations of 2D and 3D convolutions as well as the use of distinct cost functions such as the Sørensen-Dice coefficient, Cross Entropy, and a compound of them. 79% is the accuracy currently reported for the detection of benign and malign tumors using CT imagery performed by medical practitioners. The best iteration and mixture of parameters in this work resulted in an accuracy of 83% for tumor labelling. This study has further exposed the performance of a versatile and groundbreaking approach to deep learning framework designed for biomedical image segmentation.

In the modern digital landscape, mobile devices serve as crucial data generators.Their usage spans from simple communication to various applications such as userbehavior analysis and intelligent applications. However, privacy concerns associatedwith data collection are persistent. Deep learning technologies, specifically Convo-lutional Neural Networks, have been increasingly integrated into mobile applicationsas a promising solution. In this study, we evaluated the performance of a CNN im-plemented on iOS smartphones using the CIFAR-10 data set, comparing the model’saccuracy and execution time before and after conversion for on-device deployment.The overarching objective was not to design the most accurate model but to inves-tigate the feasibility of deploying machine learning models on-device while retain-ing their accuracy. The results revealed that both on-cloud and on-device modelsyielded high accuracy (93.3% and 93.25%, respectively). However, a significantdifference was observed in the total execution time, with the on-device model re-quiring a considerably longer duration (45.64 seconds) than the cloud-based model(4.55 seconds). This study provides insights into the performance of deep learningmodels on iOS smartphones, aiding in understanding their practical applications andlimitations.

Sewer pipelines are deteriorating due to aging and sub optimal material selections, leading to the infiltration of clean ground and rainfall water into the pipes. It is estimated that a significant portion (up to 40-50%) of the water entering wastewater treatment plants is actually clean infiltrated water. This infiltration not only contributes to unnecessary energy consumption but also poses the risk of flooding the sewer network and treatment plants. Finding these broken pipes is utmost importance but is not straight forward due to the pipes being located a few meters below ground. There exist methods of pinpointing where these leaks occur, but they are often time consuming and expensive. This thesis seeks to address the following question; Can the estimation of infiltration be accomplished solely through the temperature data obtained from discrete pump stations, or is the inclusion of precipitation data essential for achieving accurate results? Two machine learning algorithms are investigated to solve the regression problem of estimating the amount of rainfall derived infiltration. The first model is a classical linear regression model. The second model is a Convolutional neural network (CNN). Both of these models are trained on the same data set. The temperatures recorded at the stations are reliable and can be trusted. However, the data labeling process involves utilizing calculated flows to the stations during both dry and wet weather periods. This means that the labels of the data cannot be trusted to be the actual ground truth, and there exists an uncertainty in the data set. Both models manage to capture large temperature drops which indicates infiltration has occurred. The linear regression model seems to be too sensitive towards small temperature drops and predicts infiltration when there is none. The CNN model on the other hand seems to be able to capture only large temperature drops when infiltration occurs. However, both models are trained with data from only one station, this means that the models are biased towards the average temperature of that particular station, other stations may have a higher or lower average temperature. When testing the models on a different station with lower average temperature the models predict infiltration when there is none.

This thesis concerns the application of different artificial neural network architectures on the classification of multivariate accelerometer time series data into activity classes such as sitting, lying down, running, or walking. There is a strong correlation between increased health risks in children and their amount of daily screen time (as reported in questionnaires). The dependency is not clearly understood, as there are no such dependencies reported when the sedentary (idle) time is measured objectively. Consequently, there is an interest from the medical side to be able to perform such objective measurements. To enable large studies the measurement equipment should ideally be low-cost and non-intrusive. The report investigates how well these movement patterns can be distinguished given a certain measurement setup and a certain network structure, and how well the networks generalise to noisier data. Recurrent neural networks are given extra attention among the different networks, since they are considered well suited for data of sequential nature. Close to state-of-the-art results (95% weighted F1-score) are obtained for the tasks with 4 and 5 classes, which is notable since a considerably smaller number of sensors is used than in the previously published results. Another contribution of this thesis is that a new labeled dataset with 12 activity categories is provided, consisting of around 6 hours of recordings, comparable in number of samples to benchmarking datasets. The data collection was made in collaboration with the Department of Public Health at Karolinska Institutet. / Inom ramen för uppsatsen testas hur väl rörelsemönster kan urskiljas ur accelerometerdatamed hjälp av den gren av maskininlärning som kallas djupinlärning; där djupa artificiellaneurala nätverk av noder funktionsapproximerar mappandes från domänen av sensordatatill olika fördefinerade kategorier av aktiviteter så som gång, stående, sittande eller liggande.Det finns ett intresse från den medicinska sidan att kunna mäta fysisk aktivitet objektivt,bland annat eftersom det visats att det finns en korrelation mellan ökade hälsorisker hosbarn och deras mängd daglig skärmtid. Denna typ av mätningar ska helst kunna göras medicke-invasiv utrustning till låg kostnad för att kunna göra större studier.Enklare nätverksarkitekturer samt återimplementeringar av bästa möjliga teknik inomområdet Mänsklig aktivitetsigenkänning (HAR) testas både på ett benchmarkingdataset ochpå egeninhämtad data i samarbete med Institutet för Folkhälsovetenskap på Karolinska Institutetoch resultat redovisas för olika val av möjliga klassificeringar och olika antal dimensionerper mätpunkt. De uppnådda resultaten (95% F1-score) på ett 4- och 5-klass-problem ärjämförbara med de bästa tidigare publicerade resultaten för aktivitetsigenkänning, vilket äranmärkningsvärt då då betydligt färre accelerometrar har använts här än i de åsyftade studierna.Förutom klassificeringsresultaten som redovisas bidrar det här arbetet med ett nyttinhämtat och kategorimärkt dataset; KTH-KI-AA. Det är jämförbart i antal datapunkter medspridda benchmarkingdataset inom HAR-området.

No description available.

The first Metatarsophalangeal (MTP) joint is essential for foot biomechanics and weight-bearing activities. Osteoarthritis in this joint can lead to pain, discomfort, and limited mobility. In order to treat this, Episurf Medical is working to produce individualized implants based on 3D segmentations of the joint. As manual segmentations are both time- and cost-consuming, and susceptible to human errors, automatic approaches are preferred. This thesis uses U-Net and DeepEdit as deep-learning based methods for segmentation of the MTP joint, with the latter being evaluated with and without user interactions. The dataset used in this study consisted of 38 CT images, where each model was trained on 30 images, and the remaining images were used as a test set. The final models were evaluated and compared with regards to the Dice Similarity Coefficient (DSC), precision, and recall. The U-Net model achieved DSC 0.944, precision 0.961, and recall 0.929. The automatic DeepEdit approach obtained DSC of 0.861, precision of 0.842, and recall of 0.891, while the interactive DeepEdit approach resulted in DSC of 0.918, precision of 0.912, and recall of 0.928. All pairwise comparisons in terms of precision and DSC showed significant differences (p&lt;0.05), where U-Net had the highest performance, while the difference in recall was not found to be significant (p&gt;0.05) for any comparison. The lower performances of DeepEdit compared to U-Net could be due to lower spatial resolution in the segmentations. Nevertheless, DeepEdit remains a promising method, and further investigations of unexplored areas could be addressed as future work. / Den första Metatarsalphalangeal(MTP) leden är viktig för fotens biomekanik och viktbärande aktiviteter. Artros i denna led kan leda till smärta, obehag och begränsad rörlighet. För att behandla detta arbetar Episurf Medical med att producera individanpassade implantat baserat på 3D segmenteringar av leden. Då manuella segmenteringar både är tids- och kostnadskrävande, samt känsliga för mänskliga fel, föredras automatiska metoder. Denna avhandling använder U-Net och DeepEdit som djupinlärningsbaserade metoder för segm- entering av MTP leden, varav det senare utvärderas med och utan användarint- eraktion. Datasetet som användes i denna studie bestod av 38 CT bilder, där varje modell tränades på 30 bilder och de återstående användes som testdata. De slutliga modellerna utvärderades och jämfördes med avseende på Dice Similarity Coefficient (DSC), precision och recall. U-Net modellen uppnådde DSC 0.944, precision 0.961 och recall 0.929. Den automatiska DeepEdit metoden erhöll DSC 0.861, precision 0.842 och recall 0.891, medan den interaktiva DeepEdit metoden resulterade i DSC 0.918, precision 0.912 och recall 0.928. Alla parvisa jämförelser avseende precision och DSC visade signifikanta skillnader (p&lt;0.05), där U-Net hade den högsta prestandan, medan skillnaden i recall inte visade sig vara signifikant (p&gt;0.05) för någon jämförelse. Den lägre prestandan för DeepEdit jämfört med U-Net kan bero på lägre spatiell upplösning i segmenteringarna. Dock är DeepEdit fortfarande en lovande metod, och ytterligare undersökningar av outforskade områden kan tas upp som framtida arbete.

This dissertation will focus on the forecasting and classification of time series. Specifically, the forecasting problem will focus on imbalanced time series (ITS) which contain a mix of a mix of low probability extreme observations and high probability normal observations. Two approaches are proposed to improve the forecasting of ITS. In the first approach proposed in chapter 2, an ITS will be modelled as a composition of normal and extreme observations, the input predictor variables and the associated forecast output will be combined into moving blocks, and the blocks will be categorized as extreme event (EE) or normal event (NE) blocks. Imbalance will be decreased by oversampling the minority EE blocks and undersampling the majority NE blocks using modifications of block bootstrapping and synthetic minority oversampling technique (SMOTE). Convolution neural networks (CNNs) and long-short term memory (LSTMs) will be selected for forecast modelling. In the second approach described in chapter 3, which focuses on improving the forecasting accuracies LSTM models, a training strategy called Circular-Shift Circular Epoch Training (CSET), is proposed to preserve the natural ordering of observations in epochs during training without any attempt to balance the extreme and normal observations. The strategy will be universal because it could be applied to train LSTMs to forecast events in normal time series or in imbalanced time series in exactly the same manner. The CSET strategy will be formulated for both univariate and multivariate time series forecasting. The classification problem will focus on the classification event-related potential neural time series by exploiting information offered by the cone of influence (COI) of the continuous wavelet transform (CWT). The COI is a boundary that is superimposed on the wavelet scalogram to delineate the coefficients that are accurate from those that are inaccurate due to edge effects. The features derived from the inaccurate coefficients are, therefore, unreliable. It is hypothesized that the classifier performance would improve if unreliable features, which are outside the COI, are zeroed out, and the performance would improve even further if those features are cropped out completely. Two CNN multidomain models will be introduced to fuse the multichannel Z-scalograms and the V-scalograms. In the first multidomain model, referred to as the Z-CuboidNet, the input to the CNN will be generated by fusing the Z-scalograms of the multichannel ERPs into a frequency-time-spatial cuboid. In the second multidomain model, referred to as the V-MatrixNet, the CNN input will be formed by fusing the frequency-time vectors of the V-scalograms of the multichannel ERPs into a frequency-time-spatial matrix.