Theses (Electrical and Computer Engineering)

Permanent URI for this collection


Recent Submissions

Now showing 1 - 20 of 622
  • Item
    Infrared-Visible Image Fusion in the Gradient Domain
    (2024) Premaratne, Sanduni; Agathoklis, Panajotis; Bruton, Leonard T.
    Due to the complementary properties of the infrared cameras compared to conventional visible imaging cameras, it has become increasingly popular to fuse infrared and visible images of the same scene for better visual understanding. One major application of this is surveillance which involves videos and requires fast processing. Therefore, there is a need for investigating novel low-complexity fusion algorithms that can be implemented in real-time applications. In this study, we address this critical research problem by two-scale fusion in the gradient domain with saliency detection and image enhancement. In the proposed method, the source images are first decomposed in to base and detail layers. Next, the base parts are fused in the gradient domain by choosing the maximum absolute gradient, whereas the gradients of the detail parts are fused using a weighted average where the weights are calculated using saliency maps. Prior to fusion, the detail parts are enhanced using a guided filter-based enhancement approach. Finally, the fused gradients of the base and detail components are added together to obtain the gradients of the fused image, from which the fused image is reconstructed using a reconstruction technique based on wavelets. Experimental results demonstrate that the proposed method achieves very competitive performance in subjective and objective fusion assessments, while also outperforming most methods in terms of computational complexity.
  • Item
    Single-Class Instance Segmentation for Vectorization of Line Drawings
    (2024) Vohra, Rhythm; Branzan Albu, Alexandra
    Images can be represented and stored either in raster or in vector formats. Raster images are the most ubiquitous and are defined as matrices of pixel intensities/colours, while vector images consist of a finite set of geometric primitives, such as lines, curves, and polygons. Since geometric shapes are expressed via mathematical equations and defined by a limited number of control points, they can be manipulated in a much easier way than by directly working with pixels; hence, the vector format is much preferred to raster for image editing and understanding purposes. The conversion of a raster image into its vector correspondent is a non-trivial process, called image vectorization. Creating vector images from a given raster image can be time-consuming and requires the expertise of a skilled graphic user. This thesis explores the effectiveness of a Deep Learning based framework to vectorize raster images comprising line drawings with minimal user interventions. To improve the visual representation of the image, each stroke in the line drawing is represented with a different label and vectorized. In this document, we present an in-depth study of image vectorization, the objective of our research, challenges, potential resolutions, and compare the outcomes of our approach on six datasets consisting of different types of hand drawings. More specifically, this thesis begins by comparing raster images with vector images, the importance of image vectorization, and our objective to convert raster images to vector-based representations by accurately separating each stroke from the line drawings. In further chapters of this thesis, a Deep Learning based segmentation methodology is introduced to perform Single-Class Instance Segmentation of hand drawings to process the input raster image by labeling each pixel as belonging to a particular stroke instance. This segmentation approach is able to leverage the spatial relationships between each stroke instance. A novel loss function specifically designed to optimize our highly imbalanced datasets by scaling the margins and adding a regularization term to improve its feature selection technique. The weighted combination of our proposed margin regularized loss function is combined with the Dice loss to reduce the spatial overlap and improve the predictions over infrequent labels. Finally, the effectiveness of our segmentation technique of line drawing vectorization is compared experimentally with the state-of-the-art and our reference method. Our method can successfully handle a wide variety of human drawing styles. The results are comparable in terms of accuracy and way ahead in terms of speed and complexity, with other methods.
  • Item
    Hardware Architecture for Accelerating Frequency-Domain Ultrasound Image Reconstruction
    (2024-02-09) Navaeilavasani, Pooriya; Rakhmatov, Daler N.
    Ultrasound is a widely employed biomedical imaging modality enabling non-invasive, low-cost, and real-time diagnostics. In a typical ultrasound system, a multi-channel transducer emits sound waves into the medium and then records returning echo signals that are subsequently converted into an image of the subsurface structure. Coherent plane-wave compounding (CPWC) is one of the latest ultrasound imaging techniques that involves emitting multiple plane-wave pulses at various angles and then combining angle-specific reconstructed image data into a final frame. This approach offers high data acquisition rates (e.g., hundreds or even thousands of raw data frames per second) that are crucial for capturing fast-changing phenomena in the imaged medium. High data acquisition rates should be matched with fast data processing to increase the frame rate of reconstructed, or beamformed, image frames. One example of highly efficient plane-wave beamforming methods is the Temme-Mueller algorithm that operates in the Fourier domain. This thesis describes a novel pipelined hardware architecture for accelerating the execution of this algorithm. The proposed design has been coded in VHDL and implemented on a modern Xilinx® field-programmable gate array (FPGA), taking advantage of Xilinx® intellectual property (IP) core reuse to reduce development time. Our architecture is capable of producing over 1,300 beamformed frames per second, where each frame contains 256K complex-valued data points using the 32-bit floating-point representation for both real and imaginary parts. The correctness of our FPGA-based beamformer has been verified by comparing its output to the reference software-based implementation of the Temme-Mueller algorithm. This verification was done on an experimental ultrasound dataset available as part of the public-domain PICMUS evaluation framework. Our evaluation results demonstrate that the proposed design provides a promising alternative to the conventional GPU-based approach to high-frame-rate ultrasound image reconstruction, paving the way for future algorithmic and architectural enhancements.
  • Item
    Learning-based Ultra-Wideband Indoor Ranging and NLOS Identification
    (2024-02-09) Li, Xin; Dong, Xiaodai
    The need for precise indoor positioning has become increasingly important with the rise of Internet of Things (IoT) technology, robotics, and autonomous vehicles. Indoor positioning has a wide range of applications, including asset tracking, indoor navigation, and location-based services. To achieve high positioning precision for these applications, accurate and reliable indoor ranging is a key factor when using techniques like time of arrival (ToA), as it enables the calculation of distances between different objects in the indoor environment. In this thesis, we focus on machine learning-based approaches for indoor ranging and non-line-of-sight (NLOS) identification. The first part of the thesis concentrates on reducing ranging errors through machine learning with the improvement of the resolution of channel impulse response (CIR) data. We collect a dataset of 412, 172 traces of CIR data across 12 indoor Line-of-Sight (LOS) scenarios. This dataset is used to train and test three machine learning models, including long short-term memory (LSTM), gated recurrent units (GRU), and multi-layer perception (MLP), to predict the range between the anchor and tag directly through the CIR data. The results demonstrate that LSTM and GRU models outperform traditional meth-ods and the device built-in algorithm in terms of mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE), thereby showing the effectiveness of machine learning techniques for indoor ranging applications. On the other hand, indoor ranging accuracy can be significantly affected by NLOS conditions, where the direct path between the transmitter and receiver is obstructed, and the signal has to travel through multiple reflections and diffractions before reaching the receiver. In this thesis, we propose a quantitative approach to differentiate between Soft and Hard NLOS based on the ranging error percentage. We develop machine learning models to identify and classify NLOS conditions. Our study shows that when NLOS is classified into Soft NLOS and Hard NLOS, the accuracy of LOS identification is achieved better than using binary classification. Compared to traditional methods such as leading edge detection or search back window for ranging and positioning, our method exhibits superior performance in noise, multipath, and NLOS environments.
  • Item
    Log Message Anomaly Detection using Positive and Unlabeled Learning
    (2024-01-29) Seifishahpar, Fatemeh; Gulliver, T. Aaron
    Log messages are widely used in cloud servers and software systems. Anomaly detection of log messages is important as millions of logs are generated each day. However, besides having a complex and unstructured form, log messages are large unlabeled datasets which makes classification very difficult. In this thesis, a log message anomaly detection technique is proposed which employs Positive and Unlabeled Learning (PU Learning) to detect anomalies. Aggregated reliable negative logs are selected using the Isolation Forest, PU Learning, and Random Forest algorithms. Then, anomaly detection is conducted using deep learning Long Short-Term Memory (LSTM) network. The proposed model is evaluated using the commonly employed Openstack, BGL, and Thunderbird datasets and the results obtained indicate that the proposed model performs better than several well-known approaches in the literature.
  • Item
    StretchVADER – A Rule-based Technique to Improve Sentiment Intensity Detection using Stretched Words and Fine-Grained Sentiment Analysis
    (2024-01-22) Jokhio, Muhammad Naveed; Gulliver, Thomas Aaron
    Watching a horror movie and someone shouts “HEEEELLLPPPPPPPPP” or someone replies to your joke with a huge “HAHAHAHAHAHAHAHAHAHAHA” is known as word stretching. Word stretching is not only an integral part of spoken language but is also found in many texts. Though it is very rare in formal writing, it is frequently used on social media. Word stretching emphasizes the meaning of the underlying word, changes the context and impacts the sentiment intensity of the sentence. In this work, a rule-based fine-grained approach to sentiment analysis named StretchVADER is introduced that extends the capabilities of the rule-based approach called VADER. StretchVADER detects sentiment intensity using textual features such as stretched words and smileys by calculating a StretchVADER Score (SVS). This score is also used to label the dataset. It has been observed that many tweets contain stretched words and smileys, e.g. 28.5% in a randomly extracted dataset from Twitter. A dataset is also generated and annotated using SVS which contains detailed features related to stretched words and smileys. Finally, Machine Learning (ML) models are evaluated using two different data encoding techniques, e.g. TF-IDF and Word2Vec. The results obtained show that the XGBoost algorithm with 1500 gradient-boosted trees and TF-IDF data encoding achieved a higher accuracy, precision, recall and F1-score than the other ML models, i.e. 91.24%, 91.11%, 91.24% and 91.08%, respectively.
  • Item
    Physical Layer Authentication for Wireless Applications
    (2023-12-13) Hammouda, Mohammed; Gulliver, T. Aaron
    Internet of things (IoT) devices have become ubiquitous and go far beyond smartphones and similar devices. The IoT allows for numerous applications such as smart homes, intelligent healthcare, and intelligent transportation. However, high deployment costs limit cellular network coverage in remote and rural areas, and the reliability of cellular infrastructure during natural disasters is a concern. Thus, space and ground network integration has been proposed to provide global connectivity and support a wide range of IoT applications. Unfortunately, spoofing attacks are problematic due to network complexity and heterogeneity. Authentication for access control is an efficient way to ensure user legitimacy. However, upper layer authentication (ULA) is challenging due to limited computational power, high complexity, and communication overhead. Thus, physical layer authentication (PLA) has been proposed to aid ULA in solving these problems. PLA exploits the fact that legitimate parties and attackers have distinct physical characteristics which are unique between every pair of connected peers based on their spatial locations. In this dissertation, PLA schemes are presented using wireless attributes. First, an adaptive PLA scheme for IoT applications in urban environments is proposed using machine learning (ML) with antenna diversity to increase the number of features. A one-class classifier support vector machine (OCC-SVM) is employed using the magnitude and real and imaginary parts of the received signal at each receive antenna as features. The sounding reference signal (SRS) in the 5G uplink radio frame is employed for this purpose. Results are presented which show that this scheme provides a high authentication rate (AR) with sufficient antenna diversity. Furthermore, an adaptive PLA scheme is presented for collaboration between distributed IoT devices in multiple-input-multiple-output (MIMO) systems. The performance is evaluated considering two majority voting schemes for practical IoT applications. These schemes may be preferable for IoT devices with limited computing capabilities. An adaptive PLA scheme for low earth orbit (LEO) satellites is proposed that employs ML with Doppler frequency shift (DS) and received power (RP) features. This scheme is evaluated for fixed and mobile satellite services at different altitudes. Results are presented which show that the proposed scheme provides better authentication performance using DS and RP features together compared to using them separately. Moreover, PLA using a hypothesis test with threshold or ML for satellite authentication is presented. The results show that the AR with DS is higher than with RP at low elevation angles for both schemes, but is higher with RP at high elevation angles. Further, the ML authentication scheme provides a higher AR than the threshold scheme for a small percentage of the training data considered as outliers, but at larger percentages the OR threshold scheme is better. Finally, game-theoretic satellite authentication using physical characteristics for spoofing detection is presented. Results are given to demonstrate the effectiveness of the proposed approach.
  • Item
    Comparative Analysis of Point Sampling Strategies in Point-based 3D Object Detection
    (2023-12-11) Zhu, Rui; Li, Kin Fun
    Point-based 3D object detection has been receiving increasing attention as it can preserve the geometric information of a point cloud and avoid quantization errors or information loss caused by voxelization or projection. Point sampling plays an important role in point-based 3D detectors yet has not been thoroughly explored. In this research, we conduct a comparative analysis of three point sampling strategies to gain a deep understanding of the effect that each strategy imposes on the final performance and intermediate stages of the network. We introduce density-aware sampling and semantic-aware sampling strategies and fit them into the backbone of a lightweight and effective baseline model, aiming to reduce the density imbalance of the point cloud and better utilize semantic information. The density-aware strategy effectively balances the density but the inference time is not applicable for real-time application. Semantic-aware sampling biased on foreground points achieves a 0.19\% improvement on the baseline. Analysis on statistics and visualization reveals future research direction. We build our models on MMDetection3D platform and evaluate the performance on KITTI dataset.
  • Item
    An Accurate and Fast Animal Species Detection System for Embedded Devices
    (2023-12-06) Ibraheam, Mai Mahmoud; Li, Kin Fun; Gebali, Fayez
    Object detection is one of the vital and challenging tasks in the field of computer vision. It supports a wide range of applications in real life, such as surveillance, autonomous driving, and medical diagnostics. Object detection techniques aim to identify and localize objects of certain target classes within an image and assign each object to a corresponding class label. These techniques vary in their network architecture, training strategy, and optimization function. In this dissertation, an investigation into object detection is presented, with a specific emphasis on animal species detection. The research aims to mitigate the negative impacts of wildlife-human conflicts (WHCs) and wildlife-vehicle collisions (WVCs), particularly in remote wilderness regions/trails, urban areas/backyards, and on highways. Our goal is to enhance the accuracy and speed of animal species detection to ensure safer environments for both humans and wildlife. The research involves a comprehensive analysis of object detection techniques based on R-CNN models. Four different R-CNN models and a deformable convolutional neural network are applied on three wildlife datasets, and results are evaluated using four metrics. This comprehensive analysis informs the proposal of a novel animal species detection system. The results illustrate the system's high accuracy in distinguishing between different object categories such as animals, humans, and vehicles, as well as in identifying specific animal species. This work aims to develop an automated labelling and annotation system that eliminates the need for human intervention, thereby saving time and costs. Furthermore, it seeks to contribute to the development of robust and reliable systems which can be applied to various aspects of biological sciences, such as wildlife monitoring, conservation, and management. A key proposal of the research is to develop WHCs and WVCs real-time mitigation systems based on a lightweight animal species detection model (M-YOLO) derived from YOLOv2. Multi-level features merging is employed by adding a new pass-through layer to improve the feature extraction ability and accuracy of YOLOv2. Moreover, the two repeated 3 × 3 convolutional layers in the seventh block of the YOLOv2 architecture are removed to reduce computational complexity, and thus increase detection speed without reducing accuracy. Animal species detection methods based on regular Convolutional Neural Networks (CNNs) have been widely applied; however, these methods are difficult to adapt to geometric variations of animals in images. Thus, a modified YOLOv2 with the addition of deformable convolutional layers (DCLs) was proposed to resolve this issue. Our experimental results show that the proposed model outperforms the original YOLOv2 by 5.0% in accuracy and 12.0% in speed. Furthermore, our analysis shows that the proposed model is more suitable for deployment on embedded devices than YOLOv3 and YOLOv4. To further enhance the M-YOLO model and achieve real-time alerts on low-power and resource-constrained devices, the research proposes the integration of two key ideas: the Motion-selective Control Frames (MCF) algorithm and a parallel processing technique. These enhancements aim to minimize the detection processing delay and power consumption, which are crucial for the efficient operation of low-power, computationally limited embedded devices. Importantly, these improvements are achieved while maintaining detection accuracy.
  • Item
    Detection of Dementia: Using Electroencephalography and Machine Learning
    (2023-12-05) Ahmed, Tanveer; Gebali, Fayez; El Miligi, Haytham
    Dementia is a general term used to describe a decline in mental ability that interferes with daily life. This thesis aims to investigate the use of EEG (Electroencephalography) signals to detect dementia, which offers a promising approach in individuals with dementia, as they provide a non-invasive measure of brain activity during language tasks, which can be analyzed using machine learning algorithms to identify patterns. We also implemented various EEG features extraction and selection techniques and machine learning algorithms that have been used and provide an analysis of the results obtained. We also reported that the most people in the age bracket of 60-69 are most likely to have dementia, with females in common. Overall, K-means achieved the highest Silhouette Score for our clustering results is approximately 0.295. And Decision Tree and Random Forest models achieved the best accuracy of 95.83%. The SVM and Logistic Regression models also achieved good accuracy of 91.67% with the Decision Tree and Random Forest slightly outperforming them.
  • Item
    Authentication of Configuration Updates for Remote Field Programmable Gate Arrays with the use of Physical Unclonable Function
    (2023-11-14) Salem, Fares; Gebali, Fayez; El-Kharashi, M. Watheq
    Security has become a significant concern in our lives with the proliferation of connected devices and the emergence of embedded systems where the attack surface is getting wider every day. New security challenges and privacy concerns require secure solutions and designs. As a result, security across all computing layers, from the applications running on the devices down to the silicon, is required. Remote devices, including the ones that use field programmable gate arrays (FPGA), could be found in different industries, and require updates to be delivered securely over the air. The authentication of the remotely accessed devices is an essential requirement for the security of these systems to guarantee that the updates get delivered to trusted and authentic devices. This work proposes a secure authentication protocol for providing secure bitstream configuration updates to remote system-on-chip (SoC) FPGA devices. This is enabled using public-key cryptography and physical unclonable functions (PUF) embedded in the FPGAs that are used as the system’s hardware root-of-trust (HRoT), enabling on-demand secure cryptographic key generation, authentication, and secure session key exchange.
  • Item
    Space-Air-Ground FSO Transmission System Design for Reliable and Very-High-Throughput Satellite Communications
    (2023-11-06) Zaghloul, Ramy; Yang, Hong-Chuan
    Recent advances demonstrate satellite communications (SatComs) as a potent enabler for future Terabits/s wireless networks and bridging the digital divide. Existing SatCom systems, however, are mostly dependent on radio frequency (RF) with limited available bandwidth, which is the main bottleneck for further data rate increases. Free-space optical (FSO) communications, with huge license-free bandwidth, have emerged as a candidate alternative. Despite their ability to deliver high-throughput transmissions, FSO communications are weather-dependent and susceptible to atmospheric turbulence effects. Therefore, improving the FSO-link usability is crucial for better exploring its higher transmission rate. The goal of this research is to develop effective FSO-based SatCom solutions that can achieve very high throughput while enjoying high reliability and, as such, provide trustworthy support for future global connectivity. In this thesis, we adopt novel space-air-ground (SAG) FSO transmission approaches in the design of advanced SatCom systems for next-generation wireless networks. We first propose a reliable Terabits satellite feeder link solution. In particular, we propose a new SAG-FSO network with a strategically deployed high-altitude platform (HAP) relay to successfully remedy the atmospheric turbulence effects. We show that such a design can substantially mitigate the effects of atmospheric turbulence. Then, we integrate the proposed SAG-FSO network and hybrid single-hop (SH) FSO/RF transmission to create a SatCom feeder link with significantly improved performance and reliability. The numerical results show that the integrated transmission system achieves about 10 dB performance gain over existing solutions for both downlink and uplink scenarios. To mitigate weather effects and increase the reliability of SAG-FSO networks even further, we propose to combine SAG-FSO transmission with site diversity. With the recent technological advancements in solar cells and batteries, HAP-based relays can operate continuously for several months. A SatCom system with multiple HAP relays can enable a much more flexible design than a conventional one with multiple ground station sites. More precisely, we consider switching-based HAP relays with a hybrid SAG-FSO/RF transmission for SatCom. Our proposed system switches between HAPs based on the ground-HAP channel quality, as there are more atmospheric turbulence and weather effects. Meanwhile, the ground-HAP links corresponding to different HAP relays may experience correlated atmospheric turbulence. The obtained results illustrate that, despite the correlation adversely affecting performance, the transmission system still maintains a considerable gain over hybrid FSO/RF and single HAP systems. To increase the transmission rate for end users in a particular hot-spot area with higher traffic demand while maintaining ubiquitous coverage, we propose parallel RF and FSO transmissions to explore their complementary properties in beamwidth and bandwidth. In particular, RF transmissions serve the users over a large geographical area, while the FSO link is employed to increase the throughput to a particular hot-spot area with higher capacity demand through an access point. Independent data streams are adaptively sent over both links to satisfy capacity and availability requirements. Such a transmission strategy can effectively provide a high-speed connection to a centralized location. In addition, it can maintain ubiquitous coverage for numerous Internet-of-Things devices dispersed over a large geographical area via an RF link. We adopt the analytical system performance evaluation approach and develop efficient analytical expressions for important performance metrics for the proposed SatCom systems. Selected numerical examples and their discussions provide useful insights for engineering applications.
  • Item
    SeniorSentry: Safeguarding AgeTech Devices and Sensors Using Contextual Anomaly Detection and Supervised Machine Learning
    (2023-11-01) Nandikotkur, Achyuth; Traoré, Issa
    With the ever-growing reliance on IoT-enabled sensors to age in place, a need arises to protect them from malicious activities by detecting attacks or other anomalies. In this work, we first confirm the presence of correlations between co-located sensors by statistically analyzing two public smart-home datasets and a dataset we collected from our lab. Then, we leverage the sliding window approach and supervised machine learning to develop a novel contextual-anomaly-detection model that reaches a true positive rate of 89.47% and a false positive rate of 0%. Furthermore, as homes become smarter with these IoT sensors, the underlying communication technology they employ becomes a target for attackers. Typically, these sensors are paired with a micro-controller that has an inbuilt communication module (e.g., Bluetooth/WiFi), to form an edge device that facilitates communication. Monitoring vitals, climate control, illumination control, fall detection, incontinence detection, pill dispensing, and several other functions are successfully addressed by these devices. The family of vulnerabilities recently found in the the Link Manager Protocol (LMP) and baseband layers of the Bluetooth Classic (BT Classic) stack called BrakTooth, poses a genuine threat to the availability of such devices. In response, our research introduces a cost-effective experimental active sniffer that captures traffic at both these layers of the BT Classic stack and utilizes supervised machine learning to detect Braktooth-based attacks.
  • Item
    Efficient Federated Learning over Heterogeneous Machines
    (2023-08-31) Zhao, Lei; Lu, Wu-Sheng; Cai, Lin
    Decentralized data scattered among large number of different machines, therefore, there are two major challenges in large-scale machine learning, i.e., vast amount of data transmission over the communication network causing huge communication cost, and security and privacy concerns of the individual data. Federated Learning (FL) is a promising solution by enabling high-quality model training while keeping the data locally. We are dedicated to designing efficient FL schemes that benefit all participants. When engaging in collaborative learning with other machines, each machine must prioritize the interests of its owner. How to effectively collaborate among heterogeneous machines with diverse requirements to achieve powerful training results while maximizing the profits of each participant poses an interesting challenge in federated learning. The global model in FL attempts to acquire knowledge from each machine, while aggregating all local models into one optimal solution for every machine proves to be extremely challenging due to the heterogeneous demands. From an optimization standpoint, it becomes evident that each machine possesses its own objective function. As a result, the challenge arises as how to leverage the knowledge of others while obtaining the best possible solution for each machine. Our objective is to improve the efficiency of the collaboration of heterogeneous. The main challenges including the slow convergence speed and heavy communication cost which are caused by the heterogeneous local data sets, heterogeneous local feature spaces, and dynamic client participation. Our contributions are as follows. Firstly, the proposed transform-domain FL schemes based on Discrete Cosine Transform (DCT-FA) and Discrete Wavelet Transform (DWT-FA) aim to improve training efficiency and reduce communication burden for various IoT intelligence applications. The combination of frequency-domain and time-domain features approach (CDCT-FA) achieves higher test accuracy by combining time-domain and frequency-domain features. Secondly, the proposed collaborative learning framework for healthcare IoT devices with heterogeneous local feature spaces allows for privacy preservation while achieving performance similar to centralized training. The proposed central accelerated Federated Stochastic Variance Reduced Gradient (FSVRG) approach which in conjunction with a stochastic or deterministic client selection mechanism is shown to yield improved computational and communication efficiency contributing to the advancement of FL techniques by improving convergence speed with higher model accuracy. Additionally, the proposed data trading market and the integration of histogram of oriented gradients (HoG) with DWT approach address the challenges of data trading and non-stationarity of revenue patterns associated with data products.
  • Item
    Novel Adaptive Transmission for Effective URLLC Support in 5G and Beyond Wireless Systems: Reinforcement Learning based Designs
    (2023-08-31) Saatchi, Negin Sadat; Yang, Hong-Chuan
    The Industrial Internet of Things (IIoT) has transformed industrial processes by connecting devices and enabling real-time data exchange. However, the increasing demands of future IIoT applications necessitate a trustworthy, ultra-reliable, and low-latency communication (URLLC) service to support critical and time-sensitive operations. This requires the development of advanced wireless technologies capable of delivering data reliably while meeting stringent latency requirements. In this work, we first propose a novel adaptive transmission design for the fifthgeneration New Radio (5G NR) technology to enhance its URLLC provision capability. Our approach involves jointly selecting numerology, mini-slot size, and modulation and coding scheme (MCS) for each transmission attempt. By considering the prevailing channel conditions and the available latency budget, we aim to maximize the probability of successful data delivery while strictly adhering to latency constraints. We formulate the problem as a sequential decision-making process, which we cast as a finite-step Markov Decision Process (MDP). Our objective is to derive an optimal policy that guides the selection of transmission parameters at each step, ensuring efficient resource allocation and adaptive decision-making. To achieve this, we apply a model-based reinforcement learning and model-free deep reinforcement learning techniques to obtain the optimal policy. Through selected numerical examples, we demonstrate the superior performance of our proposed joint design compared to conventional schemes. The numerical results highlight the significant performance gains achieved across a wide range of transmission scenarios, particularly in situations with stringent latency budgets and poor channel quality. While our proposed joint design is demonstrated within the context of 5G NR, its applicability extends to future generations of wireless systems that adopt similar reliability and latency mechanisms.
  • Item
    Supervised and Unsupervised Deep Learning Methods for Underwater Image Enhancement
    (2023-08-31) Rico Espinosa, Alejandro; Branzan Albu, Alexandra
    Large amounts of underwater imagery are constantly collected for environmental monitoring studies, as they are essential for estimating marine biodiversity and abundance. However, this collected data has variable quality due to uncontrolled environmental factors that cause blur and color casting. We attempt to address this issue by proposing two novel methods for underwater image enhancement. The first part of the thesis presents a deep learning architecture that integrates elements from classical methods to simultaneously address blurriness and color casting on underwater imagery in real time. We use two parallel architectures trained in a generative adversarial network scheme (GAN) with channel and spatial attention blocks to retrieve color, and discrete wavelength transform to preserve high-frequency components. Our experiments show that our method outperforms the state-of-the-art related works with respect to the structured similarity index metric (SSIM). Qualitative comparisons with color-checkers also show notable improvements over related works. The second part of the thesis proposes an unsupervised deep-learning approach for underwater image enhancement, which eliminates the need for reference images for training. This is an important step forward as for real (not synthetic) underwater images there is no high-quality reference available. Our method is based on a mathematical model for image dehazing. We use three networks to estimate the transmission map, the atmospheric light, and the enhanced image and propose a new compound loss function. We achieve results comparable to state-of-the-art supervised methods with respect to the SSIM while performing optimally at near real-time inference speeds.
  • Item
    Improving surface plasmon resonance sensor and nanoaperture optical tweezers for biomolecule analysis
    (2023-08-30) Babaei, Elham; Gordon, Reuven
    This thesis explores innovative approaches to improve sensitivity of surface plasmon resonance (SPR) sensors and nanoaperture optical tweezers (NOT) technique for biomolecule analysis. A significant enhancement in the sensitivity of surface plasmon sensors by 3.3 times and a quadrupling of resolution are presented compared to conventional SPR sensors. The optimal design parameters for generating short-range modes on a gold film (period: 250 nm, gap size: 40 nm, thickness: 10 nm) are identified using rigorous coupled wave analysis (RCWA) to achieve minima for incident angle and wavelength, following the same configuration as conventional SPR sensors which employ a standard 50 nm thick gold film and Kretschmann-Raether coupling with a light wavelength of 760 nm. Finite difference time domain simulations confirm the correspondence of short-range surface plasmon modes to localized surface plasmons (LSP). By using the field confinement capability of short-range surface plasmon (SRSP) modes, higher sensitivity in SPR is achieved, facilitating the characterization of biomolecule interactions. The second study introduces novel approaches to enhance the colloidal lithography technique commonly used. The simplification of nanoaperture detection and characterization is achieved by charge coupled device (CCD) images and polarization-dependent transmission, eliminating the need for a scanning electron microscope (SEM), which otherwise makes the process time-consuming and costly. By employing polarization analysis, configurations of the holes, including single, trimers, and other clusters, along with their orientations, can be identified. Furthermore, changing the substrate of the sample from glass to polyvinyl chloride (PVC) results in a seven-fold decrease in the minimum required power for trapping 20 nm polystyrene beads, attributed to reduced surface repulsion. Lastly, the utilization of tape exfoliation instead of sonication in ethanol is presented, which preserves the uniform apertures on the gold surface. In the third study, an innovative method is presented for rapidly trapping single, unlabeled proteins in a NOT system. By integrating the principles of dielectrophoresis and NOT, a significant 10-fold reduction in trapping time is achieved, along with the successful trapping of Neuropeptide Y, which has a molecular weight of only 4kDa. This improvement is obtained by placing the counter electrode on a glass substrate outside the solution, thereby creating a fringe field that enhances the trapping performance. Placing the counter electrode directly in the solution does not generally lead to faster protein trapping. However, it is observed that electrophoresis can expedite the trapping of polystyrene spheres, especially with increasing applied voltage. This effect is attributed to changes in the repulsive surface potential. This voltage-dependent trapping enhancement is only observed with positive applied voltages. Furthermore, in other projects, extraordinary acoustic Raman (EAR) spectroscopy was used to trap a 20 nm polystyrene sphere. An optimal power level is observed for exciting the vibrational modes in the nanoparticle and simultaneously obtaining sharp peaks in the normalized standard deviation (NSTD) of the noise versus beat frequency spectrum. Additionally, PR65, which is a subunit of the protein phosphatase 2A (PP2A), is trapped, and its acoustic vibration modes are excited using the EAR technique. The experimental results demonstrate that the vibrational modes occur at frequencies of 9.29 GHz, 19.28 GHz, 30.23 GHz, and 41.18 GHz, which align with the results obtained through normal mode analysis (NMA). It is noteworthy that this technique offers the advantage of being single-molecule and label-free compared to conventional methods used for biomolecule characterization. The characterization of PR65 serves as a valuable model for understanding the structural regulation of various repeat proteins.
  • Item
    Local Descriptor Image Matching Acceleration and its Hardware Implementation
    (2023-08-30) Soleimani, Parastoo; Li, Kin Fun; Capson, David W.
    Computer vision algorithms have been used in an increasing number of applications during the past decades. One of the foremost challenges for using computer vision algorithms in practical applications is computational intensity which in turn may impact performance. In this dissertation, the focus is on improving speed performance by proposing novel algorithmic and hardware design techniques. Contributions are described for feature extraction and image matching. Histogram of Oriented Gradients (HOG) is one of the commonly-used algorithms for feature extraction. In order to increase the speed of computation, a hardware-software co-design is presented. The proposed design makes four contributions, including a new task allocation method which reduces resource utilization, logarithm-based bin assignment which reduces latency, parallel histogram generation for latency reduction, and a simplified block normalization technique for reducing resource utilization. The proposed design of the HOG algorithm attains comparable frame rates and is shown to use fewer hardware resources in comparison with existing work in the literature. Further contributions of this dissertation are related to the various steps of image matching algorithms, including scale-space generation, descriptor computation, and descriptor matching. For scale-space generation, a real-time FPGA-based implementation of the AKAZE algorithm with non-linear scale-space generation is proposed. The proposed implementation makes two main contributions, that include (1) mapping the two passes of the AKAZE algorithm onto a hardware architecture for parallel processing of multiple image sections, and (2) designing multi-scale line buffers for reducing resource utilization. A frame rate of 304 frames per second for a 1280×768 image resolution is achieved which is shown to be faster in comparison with other published work. For feature description, a novel circular shifting binary descriptor is proposed which leads to an efficient rotation invariant image matching. This new method eliminates complex operations such as multiplication and division from the orientation estimation step and thus significantly lowers the number of operations for descriptor computation. For descriptor matching, a novel content-addressable memory (CAM) architecture is proposed which significantly accelerates the matching step of the image matching pipeline. The time complexity of the proposed modified CAM approach to binary descriptor matching is O(n) while typically-used methods for matching have time complexity of O(n^2). The resource utilization and timing metrics for several experiments are reported to demonstrate the efficacy of the proposed design. Finally, the circular binary shifting descriptor and novel CAM matching design are applied to an experimental real-world application in aerial image matching to demonstrate the capabilities of the proposed methods.
  • Item
    Addressing Data Limitations in Defect Detection: A Case Study of Inspection in Automated Fiber Placement
    (2023-08-25) Ghamisi, Assef; Najjaran, Homayoun
    This thesis introduces novel automated visual defect detection approaches that effectively address the challenges of data scarcity and imbalance. In the manufacturing industry, conventional defect detection systems rely on end-to-end supervised learning methods that necessitate abundant labeled data, including defective samples. However, such data is often insufficiently available. In light of this, we propose two alternative approaches. The first approach combines unsupervised learning anomaly detection with rule-based computer vision, enabling effective defect detection with a smaller dataset consisting exclusively of non-defective samples. The second approach leverages rule-based computer vision exclusively, eliminating the need for any training data. To demonstrate the practicality and efficacy of the proposed approaches, this study uses the case study of Automated Fiber Placement (AFP) and design, implement, and evaluate both methods for defect detection in this industry. Specifically, these methods are tested on depth map images of the composite surface obtained using Optical Coherence Tomography (OCT) technology. Before utilizing these images for defect detection, certain preprocessing steps, such as noise filtering, are applied to enhance their quality. In the anomaly detection approach, the process begins with utilizing Hough Transform to estimate the boundaries of each composite strip (tow). Subsequently, a sliding window traverses along each tow, extracting small patches. A subset of these patches that are free from anomalies is used to train the autoencoder. Since the autoencoder is trained using normal samples, it can generate more precise reconstructions for these patches compared to abnormal ones. Consequently, the reconstruction error value serves as a quantitative metric to determine the presence of potential anomalies within each patch. By aggregating these values, an anomaly map is generated, enabling the identification of manufacturing defects within the depth map. The results demonstrate that despite the autoencoder being trained with a limited number of images, the proposed approach achieves satisfactory accuracy in binary classification and effectively localizes the defects. The rule-based method proposed in this study effectively identifies gaps and overlaps, which are the most common manufacturing defects in AFP. This approach combines classical computer vision techniques to identify the outlines of individual tows, enabling a comparison between consecutive tows to identify any potential gaps or overlaps. To assess the effectiveness of the proposed approach, this study compares the detected defects with ground truth annotations provided by human experts. The results affirm a high accuracy in segmenting gaps and overlaps.
  • Item
    Performance of OFDM and DFT-s-OFDM in the THz-Band Communications Channels
    (2023-05-18) Khorram, Erfan; Dong, Xiaodai
    The terahertz (THz) band is a promising frequency band that ranges from 300 GHz to 10 THz and being considered for the next generation of wireless networks due to the large available bandwidth and achievable ultra fast data rates. The first step in studying every wireless system is to accurately characterize the propagation channel. Therefore, in this thesis, a channel model is proposed with unique terahertz propagation characteristics that can be used to study candidate waveforms in the THz band. The next step is to study the candidate waveform designs. In this work, orthogonal frequency-division multiplexing (OFDM) and discrete Fourier transform-spread-OFDM (DFT-s-OFDM) are examined waveform candidates in the THz band. As OFDM is widely used in industry and thoroughly studied in the past couple of decades, it can be used as a benchmark for multi-carrier waveform designs. This study helps us to understand if OFDM can still be a reliable waveform in higher frequencies and how it compares with single carrier DFT-s-OFDM. In the first part of the thesis, the multi-ray communication channel is modeled based on ray tracing methods which consists of line-of-sight (LoS), reflected, and scattered paths. This model is a modified version of an existing multi-ray channel model with improvement. The coded OFDM and DFT-s-OFDM systems are studied by simulation in terms of spectral efficiency, CP length, peak-to-average power ratio (PAPR), phase noise, etc. DFT-s-OFDM is shown to possess advantages over OFDM in scattering rich THz channels with better PAPR, error performance and tolerance to phase noise, making it a preferred candidate over OFDM.