The wealth of information associated with the exponential increase in digital text, particularly within the biomedical field, has the potential to advance medical research, improve patient care, and enhance public health outcomes. However, the sheer volume and complexity of this data necessitate advanced computational tools for effective processing and analysis. We investigated the use of various pretrained transformer-based language models, particularly BERT, PubMedBERT, SciBERT, ClinicalBERT, DistilBERT, and the application of prompt engineering with GPT-4, within the context of biomedical Named Entity Recognition. Our approach incorporates a comprehensive performance evaluation analysis utilizing standard NLP evaluation metrics and computational resource usage metrics such as training time, memory usage, and inference time. Through this multifaceted approach, we sought to find out how the few-shot prompting approach using GPT4 performs in comparison to the BERT-variant language models while at the same time identifying models that not only excel in performance efficiency but also demonstrate computational affordability. Our experimental results show that even the most basic transformer-based language model outperforms the few-shot prompting approach of GPT-4, despite the popularity of the LLM in the more general Natural Language Processing tasks.

Constructing a Question Answering system is a challenging task despite a significant amount of study that has been conducted in recent times on this topic. It is even more difficult to provide satisfactory responses to the inquiries raised by users in an organizational setting as opposed to in an informal setting. We present in this paper, the results of our study into the use of a transformer-based model in the development of a technical document query system with machine reading comprehension. Our method fine-tunes a pre-trained transformer model with hyperparameter optimization using a pre-processed training dataset and tested on a different dataset. We experimented using eight pre-trained models from seven different variations of the BERT transformer architecture including BERT, RoBERTa, XLM-RoBERTa, ELECTRA, ALBERT, MobileBERT, and MPNet using the SQuAD1.1 dataset for fine-tuning and the Oracle Knowledge Documentation for testing. We found that the ALBERT pre-trained model is the best model achieving 0.891, 0.950, and 0.882 performance when measured using the Exact Match, F1 score, and Confidence Score metrics - despite its relatively small model size.

Abstract pattern is very commonly used in the textile and fashion industry. Pattern design is an area where designers need to come up with new and attractive patterns every day. It is very difficult to find employees with a sufficient creative mindset and the necessary skills to come up with new unseen attractive designs. Therefore, it would be ideal to identify a process that would allow for these patterns to be generated on their own with little to no human interaction. This can be achieved using deep learning models and techniques. One of the most recent and promising tools to solve this type of problem is Generative Adversarial Networks (GANs). In this paper, we investigate the suitability of GAN in producing abstract patterns. We achieve this by generating abstract design patterns using the two most popular GANs, namely Deep Convolutional GAN and Wasserstein GAN. By identifying the best-performing model after training using hyperparameter optimization and generating some output patterns we show that Wasserstein GAN is superior to Deep Convolutional GAN.

In autonomous driving, environment perception is an important step in understanding the driving scene. Objects in images captured through a vehicle camera can be detected and classified using semantic segmentation and depth estimation methods. Both these tasks are closely related to each other and this association helps in building a multi-task neural network where a single network is used to generate both views from a given monocular image. This approach gives the flexibility to include multiple related tasks in a single network. It helps reduce multiple independent networks and improve the performance of all related tasks. The main aim of our research presented in this paper is to build a multi-task deep learning network for simultaneous semantic segmentation and depth estimation from monocular images. Two decoder-focused U-N et-based multi-task networks that use a pre-trained Resnet-50 and DenseNet-121 which shared encoder and task-specific decoder networks with Attention Mechanisms are considered. We also employed multi-task optimization strategies such as equal weighting and dynamic weight averaging during the training of the models. The corresponding models' performance is evaluated using mean IoU for semantic segmentation and Root Mean Square Error for depth estimation. From our experiments, we found that the performance of these multi-task networks is on par with the corresponding single-task networks.

Early and accurate detection of brain tumors is very important to save the patient's life. Brain tumors are generally diagnosed manually by a radiologist by analyzing the patient”s brain MRI scans which is a time-consuming process. This led to our study of this research area for finding out a solution to automate the diagnosis to increase its speed and accuracy. In this study, we investigate the use of Residual Network deep learning architecture to diagnose and segment brain tumors. We proposed a two-step method involving a tumor detection stage, using ResNet50 architecture, and a tumor area segmentation stage using ResU-Net architecture. We adopt transfer learning on pre-trained models to help get the best performance out of the approach, as well as data augmentation to lessen the effect of data population imbalance and hyperparameter optimization to get the best set of training parameter values. Using a publicly available dataset as a testbed we show that our approach achieves 84.3 % performance outperforming the state-of-the-art using U-Net by 2% using the Dice Coefficient metric.

Automatic car damage detection and assessment are very useful in alleviating the burden of manual inspection associated with car insurance claims. This will help filter out any frivolous claims that can take up time and money to process. This problem falls into the image classification category and there has been significant progress in this field using deep learning. However, deep learning models require a large number of images for training and oftentimes this is hampered because of the lack of datasets of suitable images. This research investigates data augmentation techniques using Generative Adversarial Networks to increase the size and improve the class balance of a dataset used for training deep learning models for car damage detection and classification. We compare the performance of such an approach with one that uses a conventional data augmentation technique and with another that does not use any data augmentation. Our experiment shows that this approach has a significant improvement compared to another that does not use data augmentation and has a slight improvement compared to one that uses conventional data augmentation.

Sign Language Recognition is a form of action recognition problem. The purpose of such a system is to automatically translate sign words from one language to another. While much work has been done in the SLR domain, it is a broad area of study and numerous areas still need research attention. The work that we present in this paper aims to investigate the suitability of deep learning approaches in recognizing and classifying words from video frames in different sign languages. We consider three sign languages, namely Indian Sign Language, American Sign Language, and Turkish Sign Language. Our methodology employs five different deep learning models with increasing complexities. They are a shallow four-layer Convolutional Neural Network, a basic VGG16 model, a VGG16 model with Attention Mechanism, a VGG16 model with Transformer Encoder and Gated Recurrent Units-based Decoder, and an Inflated 3D model with the same. We trained and tested the models to recognize and classify words from videos in three different sign language datasets. From our experiment, we found that the performance of the models relates quite closely to the model's complexity with the Inflated 3D model performing the best. Furthermore, we also found that all models find it more difficult to recognize words in the American Sign Language dataset than the others.

Specialized control of pests and diseases have been a high-priority issue for the agriculture industry in many countries. On account of automation and cost effectiveness, image analytic pest recognition systems are widely utilized in practical crops prevention applications. But due to powerless hand-crafted features, current image analytic approaches achieve low accuracy and poor robustness in practical large-scale multiclass pest detection and recognition. To tackle this problem, this article proposes a novel deep learning based automatic approach using hybrid and local activated features for pest monitoring. In the presented method, we exploit the global information from feature maps to build our global activated feature pyramid network to extract pests' highly discriminative features across various scales over both depth and position levels. It makes changes of depth or spatial sensitive features in pest images more visible during downsampling. Next, an improved pest localization module named local activated region proposal network is proposed to find the precise pest objects positions by augmenting contextualized and attentional information for feature completion and enhancement in local level. The approach is evaluated on our seven-year large-scale pest data-set containing 88.6 K images (16 types of pests) with 582.1 K manually labeled pest objects. The experimental results show that our solution performs over 75.03% mean average precision (mAP) in industrial circumstances, which outweighs two other state-of-the-art methods: Faster R-CNN with mAP up to 70% and feature pyramid network mAP up to 72%.

Monitoring pest in agriculture has been a high-priority issue all over the world. Computer vision techniques are widely utilized in practical crop pest prevention applications due to the rapid development of artificial intelligence technology. However, current deep learning image analytic approaches achieve low accuracy and poor robustness in agriculture pest monitoring task. This paper targets at this challenge by proposing a novel two-stage deep learning based automatic pest monitoring system with hybrid global and local activated feature. In this approach, a Global activated Feature Pyramid Network (GaFPN) is firstly proposed for extracting highly representative features of pests over both depth and spatial position activation levels. Then, an improved Local activated Region Proposal Network (LaRPN) augmenting contextual and attentional information is represented for precisely locating pest objects. Finally, we design a fully connected neural network to estimate the severity of input image under the detected pests. The experimental results on our 88.6K images dataset (with 16 types of common pests) show that our approach outweighs the state-of-the-art methods in industrial circumstances.

This paper addresses the central problem of automatic segmentation of lumbar spine Magnetic Resonance Imaging (MRI) images to delineate boundaries between the anterior arch and posterior arch of the lumbar spine. This is necessary to efficiently detect the occurrence of lumbar spinal stenosis as a leading cause of Chronic Lower Back Pain. A patch-based classification neural network consisting of convolutional and fully connected layers is used to classify and label pixels in MRI images. The classifier is trained using overlapping patches of size 25×25 pixels taken from a set of cropped axial-view T2-weighted MRI images of the bottom three intervertebral discs. A set of experiment is conducted to measure the performance of the classification network in segmenting the images when either all or each of the discs separately is used. Using pixel accuracy, mean accuracy, mean Intersection over Union (IoU), and frequency weighted IoU as the performance metrics we have shown that our approach produces better segmentation results than eleven other pixel classifiers. Furthermore, our experiment result also indicates that our approach produces more accurate delineation of all important boundaries and making it best suited for the subsequent stage of lumbar spinal stenosis detection.

This paper examines a preliminary model guide for planning the use of Mobile Learning Technologies (MLT) in support of primary school teaching practice and delivery. A Soft Systems Methodology (SSM) approach is used in this study. The use of the SSM approach is aimed at allowing the analysis of participants' viewpoints to direct future developments in this area. Qualitative data has been collected through semi-structured interviews. The study sampled fifteen (15) primary schools from various urban and rural parts of the Phitsanulok Province in Thailand. The group of respondents consisted of seventeen (17) teachers and ten (10) school administrators, who were asked their opinion of and attitude towards the proposed use of Mobile Learning Technology (MLT) at their school. The results of this study have enabled the development of a model guide to facilitate the use of mobile learning technology to support and extend the reach of teaching in primary schools in Thailand.

The research reported in this paper examines the use of mobile learning technologies for supporting primary school teaching. Data was collected by questionnaires through an online survey in urban areas and a paper based survey in rural areas of Thailand. 398 educators and school executives participated in the research. The results of the research revealed that whilst most of the educators and school executives preferred using mobile technologies, guidance and training are necessary for teacher training in the use of such technologies. It became apparent that educators need a framework before using mobile learning technologies in order for it to be effectively utilised. The research analysed the current methods for the use of mobile technologies for supporting teaching and making learning activities enjoyable for the pupils.

Disc herniation is considered as the main cause for lower back pain (LBP), a health issue that affects a very high proportion of the UK population and is costing the UK government over £1.3 million per day in health care cost. Currently, the process to diagnose the cause of LBP involves a visual examination of a large number of Magnetic Resonance Images (MRI) but this process is both expensive in terms time and effort. Automatic detection of the lumbar disc herniation will reduce the time to diagnose and detect the cause of LBP by the orthopedist. There has been very limited progress towards automatic detection of disc herniation and all of the proposed techniques still require substantial manual intervention in many of the stages. Our analysis of the problem suggests that using the axial view of the MRI could potentially improve the outcome as opposed to the sagittal view used by these techniques. In this paper, we propose using the Centroid Distance Function as a shape feature of a segmented disc MRI taken from the axial view. Visual observation of the feature indicates that the feature could be used as a suitable indicator of the presence of herniation in the lumbar disc

Within this paper we discuss and present a novel modern 3D Geographical Information System (GIS) framework Project-Vision-Support (PVS). The framework is capable of processing large amounts of geo-spatial data to procedurally extract, extrapolate, and infer properties to create realistic real-world 3D virtual urban environments. The paper focuses on the generation of a novel scene-graph structure used in a number of algorithms and novel procedures for the increased rendering speeds of large virtual scenes and the increased processing capabilities as well as ease of use to manipulate a worlds worth of data. The scene-graph structure, made of two sections, depicts the spatial boundaries of the UKs Ordnance Survey (OS) scheme down to 1km2. Each 1km2 node contains the second section of the scene-graph structure, generated from the OpenStreetMap (OSM) classifications, involving buildings, highways, amenities, boundaries, and terrain. Leaf nodes contain the model mesh data. Generation of the spatial scene-graph for the UK takes 7.99 seconds for 6,313,150 nodes. The scene-graph structure allows for fast dispersal of render states, as well as scene manipulation by pre-categorising the data into branches of the scene-graph structure.

Digital watermarking is one of the most widely used techniques for protection of ownership rights of digital audio, images and video. Its commercial applications range from copyright protection to digital rights management. The success of a digital watermarking technology depends heavily on its robustness to withstand attacks that are aimed at removing or destroying the watermark from its host data. This paper provides analysis of a number of digital image watermark attacks and attempts to classify them into categories. A set of experimental results are also provided to show the effect of these attacks on watermarks produced using different watermarking techniques.