Engineering Commons^™

Show, Prefer And Tell: Incorporating User Preferences Into Image Captioning, Annika Lindh, Robert J. Ross, John Kelleher

Conference papers

Image Captioning (IC) is the task of generating natural language descriptions for images. Models encode the image using a convolutional neural network (CNN) and generate the caption via a recurrent model or a multi-modal transformer. Success is measured by the similarity between generated captions and human-written “ground-truth” captions, using the CIDEr [14], SPICE [1] and METEOR [2] metrics. While incremental gains have been made on these metrics, there is a lack of focus on end-user opinions on the amount of content in captions. Studies with blind and low-vision participants have found that lack of detail is a problem [6, 13, …

Detection Of Grape Clusters In Images Using Convolutional Neural Network, Mohammad Osama Shahzad, Anas Bin Aqeel, Waqar Shahid Qureshi

Articles

Convolutional Neural Networks and Deep Learning have revolutionized every field since their inception. Agriculture has also been reaping the fruits of developments in mentioned fields. Technology is being revolutionized to increase yield, save water wastage, take care of diseased weeds, and also increase the profit of farmers. Grapes are among the highest profit-yielding and important fruit related to the juice industry. Pakistan being an agricultural country, can widely benefit by cultivating and improving grapes per hectare yield. The biggest challenge in harvesting grapes to date is to detect their cluster successfully; many approaches tend to answer this problem by harvest …

Reality Analagous Synthetic Dataset Generation With Daylight Variance For Deep Learning Classification, Thomas Lee, Susan Mckeever, Jane Courtney

Conference papers

For the implementation of Autonomously navigating Unmanned Air Vehicles (UAV) in the real world, it must be shown that safe navigation is possible in all real world scenarios. In the case of UAVs powered by Deep Learning algorithms, this is a difficult task to achieve, as the weak point of any trained network is the reduction in predictive capacity when presented with unfamiliar input data. It is possible to train for more use cases, however more data is required for this, requiring time and manpower to acquire. In this work, a potential solution to the manpower issues of exponentially scaling …

Disaster Analysis Using Satellite Image Data With Knowledge Transfer And Semi-Supervised Learning Techniques, Palavi Jain

Reports

With the increase in frequency of disasters and crisis situations like floods, earthquake and hurricanes, the requirement to handle the situation efficiently through disaster response and humanitarian relief has increased. Disasters are mostly unpredictable in nature with respect to their impact on people and property. Moreover, the dynamic and varied nature of disasters makes it difficult to predict their impact accurately for advanced preparation of responses [104]. It is also notable that the economical loss due to natural disasters has increased in recent years, and it, along with the pure humanitarian need, is one of the reasons to research innovative …

Improving A Network Intrusion Detection System’S Efficiency Using Model-Based Data Augmentation, Vinicius Waterkemper Lodetti

Dissertations

A network intrusion detection system (NIDS) is one important element to mitigate cybersecurity risks, the NIDS allow for detecting anomalies in a network which may be a cyberattack to a corporate network environment. A NIDS can be seen as a classification problem where the ultimate goal is to distinguish between malicious traffic among a majority of benign traffic. Researches on NIDS are often performed using outdated datasets that don’t represent the actual cyberspace. Datasets such as the CICIDS2018 address this gap by being generated from attacks and an infrastructure that reflects an up-to-date scenario.

A problem may arise when machine …

Modern Techniques For Discovering Digital Steganography, Michael Hegarty, Anthony Keane

Conference papers

Digital steganography can be difficult to detect and as such is an ideal way of engaging in covert communications across the Internet. This research paper is a work-in-progress report on instances of steganography that were identified on websites on the Internet including some from the DarkWeb using the application of new methods of deep learning algorithms. This approach to the identification of Least Significant Bit (LSB) Steganography using Convolutional Neural Networks (CNN) has demonstrated some efficiency for image classification. The CNN algorithm was trained using datasets of images with known steganography and then applied to datasets with images to identify …

Synthetic Data Generation Using Wasserstein Conditional Gans With Gradient Penalty (Wcgans-Gp), Manhar Singh Walia

Dissertations

With data protection requirements becoming stricter, the data privacy has become increasingly important and more crucial than ever. This has led to restrictions on the availability and dissemination of real-world datasets. Synthetic data offers a viable solution to overcome barriers of data access and sharing. Existing data generation methods require a great deal of user-defined rules, manual interactions and domainspecific knowledge. Moreover, they are not able to balance the trade-off between datausability and privacy. Deep learning based methods like GANs have seen remarkable success in synthesizing images by automatically learning the complicated distributions and patterns of real data. But they …

Investigating Effect Of Amount Of Augmented Data On Performance Of Convolutional Neural Network For Multiclass Image Classification, Shivam Khandelwal

Dissertations

This research project seeks to investigate the use of Image Data augmentation that generates synthetic data by adding distortions to original images, as a means of replacement to a large amount of real data used to train the Convolutional Neural Networks. The purpose of the research project is to assess the effectiveness of augmented data over the real data by comparing the performance of the model trained with various amounts of augmented training and validation data ratio. Deep learning tasks involving convolutional neural networks have difficulty in generalizing the models effectively for computer vision tasks when the training dataset is …

Using Machine Learning Classification Methods To Detect The Presence Of Heart Disease, Nestor Pereira

Dissertations

Cardiovascular disease (CVD) is the most common cause of death in Ireland, and probably, worldwide. According to the Health Service Executive (HSE) cardiovascular disease accounting for 36% of all deaths, and one important fact, 22% of premature deaths (under age 65) are from CVD.

Using data from the Heart Disease UCI Data Set (UCI Machine Learning), we use machine learning techniques to detect the presence or absence of heart disease in the patient according to 14 features provide for this dataset. The different results are compared based on accuracy performance, confusion matrix and area under the Receiver Operating Characteristics (ROC) …

Mediaeval2019: Flood Detection In Time Sequence Satellite Images, Palavi Jain, Bianca Schoen-Phelan, Robert J. Ross

Conference papers

In this work, we present a flood detection technique from time series satellite images for the City-centered satellite sequences (CCSS) task in the MediaEval 2019 competition [1]. This work utilises a three channel feature indexing technique [13] along with a VGG16 pretrained model for automatic detection of floods. We also compared our result with RGB images and a modified NDWI technique by Mishra et al, 2015 [15]. The result shows that the three channel feature indexing technique performed the best with VGG16 and is a promising approach to detect floods from time series satellite images.

A Multi-Task Approach To Incremental Dialogue State Tracking, Anh Duong Trinh, Robert J. Ross, John D. Kelleher

Conference papers

Incrementality is a fundamental feature of language in real world use. To this point, however, the vast majority of work in automated dialogue processing has focused on language as turn based. In this paper we explore the challenge of incremental dialogue state tracking through the development and analysis of a multi-task approach to incremental dialogue state tracking. We present the design of our incremental dialogue state tracker in detail and provide evaluation against the well known Dialogue State Tracking Challenge 2 (DSTC2) dataset. In addition to a standard evaluation of the tracker, we also provide an analysis of the Incrementality …

Investigating The Application Of Deep Convolutional Neural Networks In Semi-Supervised Video Object Segmentation, Jayadeep Sasikumar

Dissertations

This thesis investigates the different approaches to video object segmentation and the current state-of-the-art in the discipline, focusing on the different deep learning techniques used to solve the problem. The primary contribution of the thesis is the investigation of usefulness of Exponential Linear Units as activation functions for deep convolutional neural architectures trained to perform object semi-supervised segmentation in videos. Mask R-CNN was chosen as the base convolutional neural architecture, with the view of extending the image segmentation algorithm to videos. Two models were created, one with Rectified Linear Units and the other with Exponential Linear Units as the respective …

Investigating The Impact Of Unsupervised Feature-Extraction From Multi-Wavelength Image Data For Photometric Classification Of Stars, Galaxies And Qsos, Annika Lindh

Conference papers

Accurate classification of astronomical objects currently relies on spectroscopic data. Acquiring this data is time-consuming and expensive compared to photometric data. Hence, improving the accuracy of photometric classification could lead to far better coverage and faster classification pipelines. This paper investigates the benefit of using unsupervised feature-extraction from multi-wavelength image data for photometric classification of stars, galaxies and QSOs. An unsupervised Deep Belief Network is used, giving the model a higher level of interpretability thanks to its generative nature and layer-wise training. A Random Forest classifier is used to measure the contribution of the novel features compared to a set …

Investigating The Impact Of Unsupervised Feature-Extraction From Multi-Wavelength Image Data For Photometric Classification Of Stars, Galaxies And Qsos, Annika Lindh

Dissertations

This thesis reviews the current state of photometric classification in Astronomy and identifies two main gaps: a dependence on handcrafted rules, and a lack of interpretability in the more successful classifiers. To address this, Deep Learning and Computer Vision were used to create a more interpretable model, using unsupervised training to reduce human bias.

The main contribution is the investigation into the impact of using unsupervised feature-extraction from multi-wavelength image data for the classification task. The feature-extraction is achieved by implementing an unsupervised Deep Belief Network to extract lower-dimensionality features from the multi-wavelength image data captured by the Sloan Digital …