SACM - United Kingdom
Permanent URI for this collectionhttps://drepo.sdl.edu.sa/handle/20.500.14154/9667
Browse
11 results
Search Results
Item Restricted Multi-Class Part Parsing based on Deep Learning(Cardiff University, 2024) Alsudays, Njuod; Wu, Jing; Lai, Yu-Kun; Ji, ZeMulti-class part parsing is a dense prediction task that seeks to simultaneously detect multiple objects and the semantic parts within these objects in the scene. This problem is important in providing detailed object understanding but is challenging due to the existence of both class-level and part-level ambiguities. This thesis investigates recent advancements in deep learning to tackle the challenges in multi-class part parsing. First, the AFPSNet network is proposed, which integrates scaled attention and feature fusion to tackle part-level ambiguity and thereby improving parts prediction accuracy. The integration of attention enhances feature representations by focusing on important features, while the feature fusion improves the fusion operation for different scales of features. An object-to-part training strategy is also used to address class-level ambiguity, improving the localisation of parts by exploiting prior knowledge of objects. Building on this foundation, the GRPSNet framework is introduced to further enhance the performance of multi-class part parsing. This framework integrates graph reasoning to capture relationships between parts, thereby improving part segmentation. These captured relationships help to enhance the recognition and localisation of parts. Moreover, the relationships of part boundaries are exploited to further enhance the accuracy of part segmentation. To further refine part segmentation, Multi-Class Boundaries integrated into the AFPSNet network. This integration aims to accurately identify and focus on the spatial boundaries of part classes, thereby enhancing the overall segmentation quality. Experimental results demonstrate the effectiveness of the proposed networks. Various evaluations, including ablation studies and comparisons with existing methods, were conducted on the widely used PASCAL-Part benchmark dataset and the large-scale ADE20K-Part benchmark dataset. These experiments validate the research hypotheses, showing notable improvements in part localisation and segmentation accuracy.24 0Item Restricted Diagnosis of Oral and maxillofacial cysts using artificial intelligence: a literature review(University of Manchester, 2024) Almohawis, Alhaitham; Yong, SinAbstract Oral and maxillofacial cysts are cavities that can pose significant risks if not detected and treated promptly. Many of these cysts are asymptomatic, often going unnoticed until complications arise. The introduction of artificial intelligence (AI) presents a promising opportunity for early detection and management of these cysts. Aim: To explore current studies on the use of artificial intelligence in diagnosing oral and maxillofacial cysts. Objectives: To examine the existing literature in this field, assess the accuracy, effectiveness, and limitations of AI models, and identify challenges in implementing AI in clinical practice. Methods: This literature review followed a systematic approach, identifying 223 studies from PUBMED and SCOPUS databases between 1975 and 2024. After applying inclusion and exclusion criteria, 26 retrospective cohort studies were included in the final analysis. A risk of bias assessment was conducted using the ROBINS I tool. Results: The investigation revealed that AI models consistently demonstrate high accuracy in detecting oral cysts in both radiographs and digital histopathology. The ROBINS I tool indicated a moderate risk of bias in most of the included studies. Notable limitations include limited datasets, variable data quality, and a lack of explainability in AI models results. Conclusion: AI models have shown considerable effectiveness and speed in detecting both simple and complex cysts. However, to fully leverage AI's potential in clinical settings, further rigorous studies are needed to evaluate its risks, benefits, and feasibility, ensuring compliance with governmental health regulations on AI.13 0Item Restricted Exploring the Applications of Artificial Intelligence in Enhancing Pre-Hospital Care: A Scoping Review(Queen’s University, Belfast, 2024) Alfaifi, Yahya; Clarke, SusanArtificial Intelligence (AI) has the potential to significantly improve pre-hospital care, especially in emergency medical services (EMS). However, its current application remains scattered, with varying integration levels across care stages. This scoping review aims to map and assess existing research on AI applications within pre-hospital care without focusing on specific AI technologies, such as machine learning (ML), deep learning (DL), or decision support systems (DSS). The review reflects the current research landscape, capturing how AI is utilised across critical stages such as call-taking, dispatch, and on-scene assessment. Using the framework developed by Arksey and O’Malley (2005), a systematic search was conducted across multiple databases to identify studies relevant to AI in pre-hospital care. The scope was deliberately broad to capture a comprehensive view of the available literature, focusing on identifying areas where further research is needed. The findings indicate that DSS is commonly used to support decision-making in call-taking and dispatch, while more advanced AI applications like ML and DL show potential in predictive analytics and real-time decision-making. However, these technologies are still in their early stages of real-world implementation. This review highlights the gaps in AI research, particularly in the later stages of prehospital care, such as transport and handover. Further exploration is necessary to unlock AI’s full potential in enhancing EMS operations and outcomes.46 0Item Restricted The Role of Artificial Intelligence in Breast Cancer Screening Programmes: A Literature Review and Focus Upon Policy Implications(The University of Edinburgh, 2024) Alrabiah, Alanoud; Hellowell, MarkBackground: Breast cancer (BC) is a leading cause of morbidity and mortality amongst older women, leading to the introduction of screening programmes to support earlier detection and improved survivability. Current screening programmes rely upon the performance of radiologists in terms of accuracy; however, evidence shows that both under and overdiagnosis means screening also results in harms to some women. Artificial intelligence is then a promising technology for improving the accuracy of mammogram screening. Aim: To describe the potential roles of AI in BC screening, and the potential benefits, limitations and risks in these roles. Methods: PubMed, SCOPUS, and CINAHL were searched. Primary research studies published in English and in the last ten years, investigating the accuracy of AI systems for screening BC, were eligible for review. Evidence was appraised using the CASP (2024) checklists and data analysed narratively. Results: 14 studies were found eligible for review, mostly adopting a retrospective study design or laboratory study design. Roles for AI in BC screening include as a standalone system replacing radiologists entirely, as risk stratification systems used before radiologist readings, or as reader aids. While some studies reported AI systems to be superior, others reported accuracy to be inferior to radiologist readings. Differences in results could be due to variations in AI system or radiologist performance. Conclusion: There is insufficient evidence to support the use of AI in BC screening programmes, and more robust, prospective studies comparing readings from clinical practice are urgently required. Policy must also be implemented to regulate the use of AI until there is sufficient evidence to support its use.12 0Item Restricted Sketch compression(University of Surre, 2023-09) Alsadoun, Hadeel Mohammed; song, Yi-Zhe; Ashcroft, AlexanderIn the rapidly evolving field of digital art and animation, traditional sketching techniques often rely on pixel-based methods, leading to less meaningful representations. This dissertation aims to transform this paradigm by rigorously investigating the efficacy of autoencoders for vector sketch compression. We conducted experiments using two distinct neural network architectures: Long Short-Term Memory (LSTM) and Transformer-based autoencoders. The Transformer model, which has significantly impacted the field of sequence-to-sequence tasks, especially in natural language processing, serves as a focal point of our study. Our experiment aims to answer a compelling question: Can these impressive results be replicated in the domain of vector sketch compression? The answer is a resounding yes. The Transformer model not only excelled in reconstructing sketches but also simplified the strokes and enhanced the overall quality of the sketch, achieving an impressive 85.03% classification accuracy. The LSTM model, known for its ability to capture temporal dependencies, served as our baseline, achieving a classification accuracy of 56.139% on a pre-trained classifier. Our findings strongly advocate for the adoption of Transformer-based models in vector sketch compression, offering a more compact and semantically rich representation. The LSTM model’s respectable performance also suggests its potential utility in less complex scenarios. Overall, this study opens new avenues for research in digital art, particularly in optimizing Transformer architectures for sketch compression.12 0Item Restricted Exploring AI-Powered Image Generation for Fashion Collaborations: A Deep Learning Approach to Blending Shoe Designs from Multiple Brands(Saudi Digital Library, 2023-09-07) AlJoudi, Sarh; Marchioni, EnricoThis thesis presents an approach for fashion shoe design generation using Generative Adversarial Networks (GANs) with a focus on the WGAN-GP, ProGAN, and StyleGAN architectures. The primary objective is to assess the visual quality, diversity, and adherence to the fashion trends of the generated shoe designs. A set of evaluation metrics, including Fréchet Inception Distance (FID), Earth Mover’s Distance (EMD), Maximum Mean Discrepancy (MMD), and K-Nearest Neighbours (KNN) accuracy, are employed to quantitatively evaluate the performance of each model. The evaluation results demonstrate that ProGAN outperforms both WGAN-GP and StyleGAN in all metrics, achieving the lowest FID, EMD, and MMD scores. ProGAN generates visually appealing and diverse shoe designs, that closely resembling real-world fashion trends. While WGAN-GP also achieves acceptable results, StyleGAN faces challenges with the droplet effect and noisy colours in the generated designs, resulting in lower performance in all evaluation metrics. Nevertheless, StyleGAN’s style-mixing capability showcases its potential for creating novel and creative shoe designs. The qualitative evaluation further confirms the ProGAN as the preferred choice for generating high-quality and fashion-forward shoe designs. Its robustness, progressive growing approach, and architectural stability contribute to its outstanding performance. The discussion also highlights the potential implications of GAN-based fashion image generation in the fashion industry and creative domains, such as trend forecasting and virtual garment display. The findings suggest that ProGAN holds promise as a reliable and creative tool for generating fashion-forward shoe designs and adopting StyleGAN 2 with the Noise-Conditional AdaIN (NCAIN) layer may address the droplet effect problem and enhance visual coherency in the generated designs. Ultimately, the results and discussions pave the way for further advancements and applications of GAN-based fashion design in the fashion industry and creative domains. Keywords: Deep learning, shoe design, GANs, style-mixing.49 0Item Restricted Universal Deep Learning for Signal Detection in Wireless Networks(Saudi Digital Library, 2023-11-25) Albagami, Khalid; Li, Geoffrey YeRecently, applying deep learning (DL) in physical layer communications for channel estimation and signal detection has extensively been studied in the literature. The majority of the existing studies investigating the use of DL techniques to address the channel estimation and signal detection problem focus on analysing channel impulse response (CIR) that are generated from only one channel type distribution such as additive white Gaussian channel Noise (AWGN) and Rayleigh channel. Although this field is well-researched in the literature, DL models are yet to be widely adopted in practice for channel estimation and signal detection. The main challenge is that a pre-trained DL model exhibits sub-optimal performance after transitioning to a different wireless channel environment. Moreover, with the limited resources and processing units available at the base station, it is a cumbersome task to re-collect data and re-train the model whenever the wireless environment changes. In this thesis, we study and investigate the feasibility of applying universal deep neural network (Uni-DNN) model consisting of two cascaded DL models where the first one works as a wireless channel classifier and the second one as a signal detector. The channel classifier DL model works on identifying the wireless channel distribution that impaired the transmitted signal. Then, the signal detector DL model utilises this information along with the received signal to recover the transmitted signal. Three different Uni-DNN architectures were developed namely architectures A, B and C utilising a combination of deep neural network (DNN) and convolutional neural network (CNN) models. The proposed architectures are trained and tested on multiple frequency selective and frequency flat fast fading wireless channels and their bit error rate (BER) performance is compared to both conventional DL models and popular channel estimation techniques such as minimum mean square error (MMSE) and least square (LS) for orthogonal frequency division multiplexing (OFDM) multiplexing scheme under AWGN with signal-to-noise ratio (SNR) ranging from 0 to 20dB. The Uni-DNN architecture C with cascaded DNN-CNN has shown the most promising results out of the three in adapting to multiple wireless channels, outperforming conventional DL models, LS and nonperfect-CSI MMSE conventional channel estimators in low pilot frequency density. In addition, the theoretical and practical Uni-DNN model inference time complexity analysis shows that such a model is faster to compute compared to MMSE and can be deployed in real-time applications.26 0Item Restricted A Comparison of Time Series and Deep Learning Methods for Predicting Stock Prices(Saudi Digital Library, 2023-03-23) Alajmi, Shahad; Ji, LanpengStock market is one of the most competitive financial markets where investors need to know the trend of prices in advance. There have been many improvements and advancements in the application of neural networks in the financial industry. In this research, two advanced methods were used to simulate and predict the close stock prices of Saudi Telecom Company (STC). The first method was the autoregressive integrated moving average (ARIMA) and the second method was a by using a class of deep learning neural networks called recurrent neural networks (RNN). ARIMA (p,d,q) was the statistical method selected as a time series model based on the level, trend ,and seasonality of data. Additionally, the order of p and q is based on an autocorrelation function ‘ACF’ ,and a partial autocorrelation function ‘PACF’, the optimal model of this research was ARIMA (1,0,28). Moreover, RNN uses long short-term memory layers (LSTMs), dropout regularisation, activation functions, a loss function, which is the mean square error (MSE), and the Adam optimiser to simulate the predictions. The unique characteristic of LSTMs is that the model is able to store previous data over time and use this data to predict future prices. The structure of the LSTM consists of five layers: one input layer, three hidden layers ,and one output layer. The methods used to measure the performance of predictions of each model are the mean absolute percentage error ‘MAPE’ and the root squared mean error ‘RMSE’. ARIMA (1,0,28) is the model that was found to have a lower error between actual and predicted prices. After analysing each model, ARIMA model’s prediction accuracy was 96.3% and RNN’s accuracy was 93.8%; we concluded that the ARIMA model is better than the RNN model for forecasting the close stock prices of STC. This research include important data which can benefit investors and companies to make economical decisions, such as when to buy or sell shares.26 0Item Restricted Machine Learning Methods for Human Identification from Dorsal Hand Images(2023-06-27) Alghamdi, Mona; Angelov, Plamen; Rahmani, HusseinPerson identification is a process that uniquely identifies an individual based on physical or behavioural traits. This study investigates methods for the analysis of images of the human hand, focusing on their uniqueness and potential use for human identification. The human hand has significant and distinctive characteristics, and is highly complex and interesting, yet it has not been explored in much detail, particularly in the context of the contemporary high level of digitalisation and, more specifically, the advances in artificial intelligence (AI), machine learning (ML) and computer vision (CV). This research area is highly multi-disciplinary, involving anatomists, anthropologists, bioinformaticians, image analysts and, increasingly, computer scientists. A growing pool of advanced methods based on AI, ML and CV can benefit and relate directly to a better representation of the human hand in computer analysis. Historically, the research methods in this area relied on ‘handcrafted’ features such as the local binary pattern (LBP) and histogram of gradient (HOG) extraction, which necessitated human intervention. However, such approaches struggle to encode the human hand in variable conditions effectively, because of the change in camera viewpoint, hand pose, rotation, image quality, and self-occlusion. Thus, their performance is limited. Recently, there has been a surge of interest in deep learning neural network (DLNN) approaches, specifically convolutional neural networks (CNNs), due to the highly accurate results achieved in many applications and the wide availability of images. This work investigates advanced methods based on ML and DLNN for segmenting hand images with various rotation changes into different patches (e.g., knuckles and fingernails). The thesis focuses on developing ML methods like pre-trained CNN models on the 'ImageNet' dataset to learn the underlying structure of the human hand by extracting robust features from hand images with diverse conditions of rotation, and image quality. Also, this study investigates fine-tuning the pre-trained models of DLNN on subsets of hand images, as well as using various similarity metrics to find the best match of the individual’s hand. Furthermore, this work explores different types of ensemble learning or fusions, those of different region and similarity metrics to improve human identification results. This thesis also presents a study of a Siamese network on sub-images or segments of human dorsal hands for identification tasks. All presented approaches are compared with the state-of-the-art methods. This study advances the understanding of variations in and the uniqueness of humans using patches of their hands (e.g., different types of knuckles and fingernails). Lastly, it compares the matching performances of the left- and right-hand patches using various hand datasets and investigates whether the fingernail produces better identification results than the knuckles. This research shows that the proposed framework for person identification based on hand components led to better person identification results. The framework consists of vital feature extractions based on deep learning neural network (DLNN) and similarity metrics. These elements enhanced the performance. Also, the fingernails' shape performed better than other hand components, including the base, major, and minor knuckles. The left hand can be more distinguishable to individuals than the right hand. The fine-tuning of the hand components and ensemble learning improved the identification results.44 0Item Restricted The Terminator: An AI-Based Framework to Handle Dependability Threats in Large-Scale Distributed Systems(University of Warwick, 2023-06-28) Alharthi, Khalid Ayed Budayai; Juhmika, ArshadWith the advent of resource-hungry applications such as scientific simulations and artificial intelligence (AI), the need for high-performance computing (HPC) infrastructure is becoming more pressing. HPC systems are typically characterised by the scale of the resources they possess, containing a large number of sophisticated HW components that are tightly integrated. This scale and design complexity inherently contribute to sources of uncertainties, i.e., there are dependability threats that perturb the system during application execution. During system execution, these HPC systems generate a massive amount of log messages that capture the health status of the various components. Several previous works have leveraged those systems’ logs for dependability purposes, such as failure prediction, with varying results. In this work, three novel AI-based techniques are proposed to address two major dependability problems, those of (i) error detection and (ii) failure prediction. The proposed error detection technique leverages the sentiments embedded in log messages in a novel way, making the approach HPC system-independent, i.e., the technique can be used to detect errors in any HPC system. On the other hand, two novel self-supervised transformer neural networks are developed for failure prediction, thereby obviating the need for labels, which are notoriously difficult to obtain in HPC systems. The first transformer technique, called Clairvoyant, accurately predicts the location of the failure, while the second technique, called Time Machine, extends Clairvoyant by also accurately predicting the lead time to failure (LTTF). Time Machine addresses the typical regression problem of LTTF as a novel multi-class classification problem, using a novel oversampling method for online time-based task training. Results from six real-world HPC clusters’ datasets show that our approaches significantly outperform the state-of-the-art methods on various metrics.19 0