Research Bank
Not a member yet
5195 research outputs found
Sort by
Speech emotion recognition using transfer learning methods
No full textRESEARCH QUESTIONS
•What are the current limitations of the existing speech-emotion recognition models?
• How data augmentation methods can be used to generate synthetic data that could improve the performance of SER systems?
• How to use the Transfer learning methods with a fine-tuning process to utilize its benefits?
ABSTRACT
Identifying human emotions using machine learning techniques in this technological advancement era is more popular, and it offers practical contributions to different fields such as health and medicine. Identifying human emotions can be made possible by analyzing their facial expressions, body gestures, and speech. With the advancement of automatic speech recognition, emotion recognition using speech is more effective and accurate than other methods. Most existing systems trained with insufficient amounts of data face the issue of performing effectively when confronted with unseen data. These systems mostly used traditional machine learning techniques.
This thesis focuses on identifying speech emotions using transfer learning, which uses pre-trained models with large-scale data. Therefore, due to the lessening of training duration and the computational cost, the efficiency of the process of speech emotion recognition will increase. Combining several datasets and using augmented data can help the existing limitation of lack of data for training
Genre prompts for professional practice creative non-fiction
No full textThis file is the data behind this paper which is accepted for publication: Guruge, D., Mann, S., Myers, R., Bates, O., Goldweber, M., Williamson, A., Lasenby, J., & Brooks, I. (2025). Surviving the narrative collapse: Sustainability and justice in computing within limits. LIMITS ’25, June 26 –27
Apple detection using filters under varying lighting conditions
No full textRESEARCH QUESTIONS
• How do different lighting conditions affect apple detection in orchard images?
• How can the lighting conditions or noise of the captured image be identified?
• What filters need to be applied to the image to get the enhanced image for object detection?
• Which image filters are most effective in enhancing apple detection under varying lighting conditions or noise levels in the image by using YOLOv9?
ABSTRACT
The advancement of intelligent farming and agricultural automation has led to the development of robotic methods for efficient apple harvesting in recent years. However, because of complicated backgrounds and fluctuating illumination, it is still challenging to identify apples in genuine orchard settings. Issues like shifting lighting intensities, shadows, and image noise complicate detection, impacting the reliability of these automated systems in real-world applications. Identifying apples under different lighting conditions, such as sunlight, shadow, darkness, and blur, presents a significant obstacle in automating apple harvesting. Traditional object detection algorithms struggle with these unpredictable lighting variations, leading to decreased accuracy. There is a need for a system that can dynamically adapt to these environmental challenges and improve apple detection reliability in varying orchard conditions.
This research investigates the integration of Support Vector Machine classification and image preprocessing techniques as a preprocessing step for YOLOv9-based object detection. By classifying images based on their lighting conditions and applying appropriate filters, the research aims to enhance image clarity before detection. The study further explores various filters, such as contrast adjustment, gamma correction, and histogram equalization, to determine their effectiveness in improving detection under specific lighting scenarios. The proposed solution involves a multi-phase approach where an SVM classifier identifies the lighting condition of each image, and the corresponding filter is applied to optimize image quality. YOLOv9, an efficient object detection model, is then used to detect apples in the pre processed images. This adaptive system enables real-time preprocessing tailored to lighting variations, ensuring that each image is processed optimally for improved feature extraction and detection accuracy.
The experimental results demonstrate that the integrated SVM-Filter-YOLOv9 pipeline significantly enhances detection rates, particularly in low-light and high-contrast environments. By employing adaptive gamma correction, contrast stretching, and colour enhancement filters, the system achieved an overall improvement of approximately 23% in apple detection accuracy compared to non-filtered images. Notably, individual filter performance was especially effective for dark conditions, showing a 47% improvement, and for sunlight conditions, with a 37% increase. This study provides a scalable and reliable framework for automated fruit detection systems, positioning it as a valuable tool for future robotic harvesting technologies
Invisible village: A potential alternative for urban village redevelopment in China
No full textRESEARCH QUESTION
How can an architectural intervention contribute to a more functional and
culturally identifiable community in an urban village redevelopment
project in Langfang, China?
ABSTRACT
This research explores an alternative architectural strategy for urban village regeneration, focusing on Dong Village in Langfang, China. The project investigates how architectural interventions can contribute to creating a more functional and culturally identifiable community in the context of rapid urbanisation and government-led redevelopment. It questions the dominant policy of demolition and reconstruction, which, although improving physical infrastructure, often eradicates social memory and local identity. It seeks to demonstrate that spatial transformation does not have to destroy the community roots but can instead support cultural continuity and a renewed sense of belonging.
The research focuses on the evolution of urban villages, examining how these communities emerged through land ownership shifts and informal development under China’s unique urban-rural structure. While previous studies have largely prioritised administrative and technical upgrading, this project identifies a critical gap in addressing the preservation of cultural identity and social networks. In response, the study positions architectural design as a tool to bridge past and present, exploring how productive landscapes, particularly greenhouse structures, can reconnect residents with their agricultural memory and enhance communal life. The research combines site study, theoretical investigation, precedent analysis, and design experimentation. Drawing on urban planning theories, such as Ebenezer Howard’s garden city vision, Kevin Lynch’s concepts of spatial legibility, and Jane Jacobs’s advocacy for social diversity, the project proposes a vibrant neighbourhood that integrates residential units, public spaces, and productive landscapes. Greenhouses are introduced as symbolic and functional nodes that promote social interaction and ecological awareness. The design engages with local spatial patterns and community narratives, offering a framework for regeneration that emphasises cultural continuity.
This project contributes a design-led perspective to the field of urban renewal, proposing adaptable strategies that challenge conventional redevelopment approaches and highlight the importance of place, memory, and indigenous community
The collar restraint vs the harness restraint: The pulling behaviour of restrained dogs during walks
No full textTwo types of restraint are commonly used by owners when walking dogs (Canis lupus familiaris) on a leash. These are the traditional neck collar and the more modern harness. The harness restraint redistributes pressure from the throat to the chest and under the front limbs when dogs pull. This broader distribution of pressure from the harness may reduce the potential for harm or discomfort when leash pulling occurs, leading to its rise in popularity in recent times as a more humane restraint method. There is also a belief that harnesses may reduce pulling by dogs, also contributing to their popularity, but evidence on this is limited. Our study compared the popularity of the two restraint types by public dog walkers, and the frequency and duration of pulling behaviours of dogs walking on a leash at three public parks in New Zealand. Eighty-nine randomly selected dogs were observed using focal animal continuous sampling for 5 minutes each. Only 40.45% of these dogs wore a harness. Dogs on a collar were more likely to pull (37.74%) compared to those on a harness (27.78%) and pulled for slightly longer (mean = 46.95s ±84.23s compared to 32.58s ±74.42s). Medium sized dogs were more likely than small dogs to be walked on a collar restraint than small dogs, while large dogs were 50/50. Interestingly, medium sized dogs pulled for longer (53.05s ±92.79s) than small (25.33s ±56.17s) or large (29.41s ±71.14s) dogs, especially when on a collar (61.79s) compared to a harness (31.19s). Our study suggests that harnesses do reduce pulling behaviour in dogs, although a larger sample size is needed to confirm. Owners with dogs that pull when being walked should consider using a harness restraint
The karoro Larus dominicanus in northern Aotearoa | New Zealand: Diet and evidence of changing trophic position from regurgitated pellets and stable isotope analysis of contemporary and historic feathers and bones
No full textCoastal seabirds are valuable indicators of ecological change in nearshore marine systems impacted by human activities. This study examined how human population growth and urban expansion have influenced the long-term dietary patterns of karoro (southern black-backed gull Larus dominicanus) in Auckland, Aotearoa | New Zealand. Specifically, we assessed whether increasing urbanisation has led to a dietary shift from marine-based prey to greater reliance on terrestrial and anthropogenic food sources. Contemporary diet composition was analysed using regurgitated pellets from two island breeding sites (Rangitoto and Tiritiri Matangi) in the Hauraki Gulf—sites differing in proximity to urban influence—and an urban roosting site at Western Springs Park. To assess long-term dietary trends, we conducted nitrogen stable isotope analysis of feathers from museum specimens spanning 109 years and collagen from contemporary karoro bones and subfossil bones from Tokerau Beach predating human arrival in Aotearoa. Pellet analysis indicated a diverse diet comprising marine vertebrates, invertebrates, and terrestrial or anthropogenic food sources, with city proximity influencing dietary composition. Stable isotope analysis revealed significant changes in karoro trophic ecology over the past century with birds shifting from a predominantly marine-based diet to one more reliant on terrestrial food, likely due to declining marine prey and increased urban food availability. Stable isotope data from ancient bones indicated that pre-human karoro were obligate coastal marine predators and scavengers, occupying a higher trophic level than their modern counterparts. These findings provide insight into how urbanisation and ecological change have shaped karoro diets over time
Work-based learning in cybersecurity
No full textWork-based learning - benefits
Challenges and solutions
Cybersecurity Diploma internship model
Cybersecurity internship commitment
Consistency Review 2024
Further improvement
Conclusio
Sepsis prediction in ICU patients using deep learning
No full textRESEARCH QUESTIONS
• How does the performance of the CNN-LSTM model compare to traditional machine learning models (e.g., Decision Tree, Random Forest, SVM) in terms of accuracy, precision, recall, and F1-score for early sepsis detection?
• What is the minimum subset of clinical features (e.g., heart rate, temperature, lactate levels) required to maintain a high-performing prediction model with reduced computational cost?
• How do SHAP (SHapley Additive exPlanations) values contribute to the interpretability of the CNN-LSTM model and support clinical decision-making in ICU environments?
• What are the limitations of using synthetic oversampling techniques such as SMOTE in the context of highly imbalanced clinical datasets, and how do they affect the model’s generalizability?
ABSTRACT
Sepsis, the most lethal cause of intensive care unit (ICU) admissions, is a life-threatening condition. Early and correct diagnosis is key in improving survival. The conventional tools, like SOFA and SIRS scores, are static and retrospective, thereby resulting into a delayed identifying of the condition, and also to false positive detection. This is a pilot study aimed to solve this set of challenging questions by a novel approach of deep learning based early sepsis prediction using multi-variate time-series lab test, vital sign, and electronic health record (EHR) data.
The proposed approach leverages Convolutional Neural Networks (CNNs) for automatic data feature extraction, combined with Long Short-Term Memory (LSTM) networks to identify temporal progressions in patients’ data. Database preprocessing by normalisation, imputation, and the Synthetic Minority Over-sampling Technique (SMOTE) is utilized to correctly handle common issues, e.g., noise, missing values, and class imbalance. Interpretability is strengthened by explanations generated via Shapley Additive explanations (SHAP), thus promoting clinical use and enabling clear interpretation of model predictions.
Our CNN-LSTM model also performed well in 22,824 test samples, with the training data consisting of almost 58,000 hourly CCU data: accuracy 95.63%, precision 97.6%, recall 93.5%, F1-score 95.5% for sepsis classification. The high NPV for excluding non-sepsis was similarly established. The attention mechanism allowed the model to be significantly more interpretable by focusing on key prediction characteristics.
Together, this work shows that deep learning can predict sepsis with high accuracy well in advance of the manifestation of clinical symptoms and may be used as a practical way to trigger early intervention and improve outcomes in the ICU
Financial fraud and accountability in ancient Sri Lanka: A forensic accounting perspective on the Kings' Period
No full textRESEARCH QUESTIONS
• RQ1: What types of financial frauds and allegations were reported during the Ancient Kings period in Sri Lanka (377 BCE–1017 CE)?
• RQ2: How were financial frauds and allegations detected and prevented during the Ancient Kings period in Sri Lanka (377 BCE–1017 CE) ?
• RQ3: How were financial fraud and allegations investigated, documented and addressed during the Ancient Kings period (377 BCE–1017 CE)
BACKGROUND OF STUDY
• A growing demand for forensic accounting and investigation due to financial scandals not only in Sri Lanka (Rathnasiri and Bandara, 2017) but also globally (Abdullahi et al., 2018; Alshurafat et al., 2020)
• Forensic accounting is the application of accounting, analytical, and report-writing skills to conduct financial investigations, assist in resolving actual or anticipated disputes, and perform valuations of businesses and other assets (Hegazy et al., 2017)
• While the definition of forensic accounting has evolved over time, its historical development remains underexplored (Golden et al., 2006; Bhasin, 2007; Okoye and Akenbor, 2009) • The history of accounting is also helpful as it provides perspectives that can influence the directions and future trends of the profession (Farag, 2009
Behaviour, binge-watching and big data: Powered by AI
No full textRESEARCH IN OUR LAB
• We have 24,000 hrs of video to systematically assess guinea pig behaviour as a function of size and shape of enclosures.
• Preliminary analysis but not achievable by humans
• Computer Science student has a created an AI learning tool to binge the videos for us!
RESEARCH AIM
• Investigate how ramp COVER influences climbing behavior
• Explore how the trade-off between safety and vigilance might affect analysis of behaviour
• Hypothesis: Full cover increases climbing speed due to perceived safet