Imaging, Data and Communications

Tarek is a medical engineer, a final-year PhD researcher, and an Engineering Teaching Assistant at the School of Engineering. He joined Edinburgh in January 2020, completing a joint MSc in Sensor and Imaging Systems between the University of Glasgow and the University of Edinburgh.

His research project is undertaken in partnership with GlaxoSmithKline (GSK) and National Physical Laboratory (NPL) and will focus on developing image processing and machine learning techniques for evaluating disease and drug effectiveness in fibre-bundle endomicroscopy systems.

• MSc in Sensor and Imaging Systems. University of Glasgow, Scotland. 2020
• BSc in Medical Engineering. Al-Ahliyya Amman University, Jordan. 2014

Tarek is also an Academic Engagement Coordinator at the Postgraduate Institute for Measurement Science (PGI). He is actively involved in creating several PGI publications and in organising and chairing the seventh annual PGI conference 2023. (For more information email: tarek.haloubi@npl.co.uk)

Research Theme

Sensor Signal Processing

Aim

Develop new computationally efficient algorithms for large multi-dimensional array processing.

Objectives

1. Develop new fast beamforming algorithms and adaptive array processing techniques using advances in fast Fourier integral operators and randomized algorithms
2. Build applications using these techniques for high-dimensional wideband sensor array processing problems, e.g. requiring the simultaneous processing of range, doppler, azimuth, elevation and frequency, in challenging acoustic and/or radar domains.
3. Explore the performance trade-off between accuracy and computational efficiency.

Description

Large scale, multi-dimensional array processing problems, e.g. simultaneously processing range, doppler, azimuth, elevation and frequency, are ubiquitous in radar and sonar sensing, and it is essential to keep the computation to a minimum in order to process the outputs in a timely manner. For narrowband uniformly spaced linear arrays fast transforms such as the FFT can be exploited. However, dealing with large wideband non-uniform arrays remains a major challenge.

This project will develop novel fast array processing algorithms for statistical estimation and beamforming based on two recent developments in signal processing and related mathematics. The first is the construction of fast Fourier integral operator approximations [1,2,3,4], that can provably approximate the Fourier integral to a given level of accuracy in O(n.log(n)) while accommodating different sampling/physical geometries. The second is efficient fast randomized algorithms for solving large least square problems [5,6] that offer the potential for novel beamforming solutions [7]. The project will explore how these ideas can be integrated to develop novel efficient array processing solutions to tackle challenging sonar and RF array processing problems and to assess their performance trade-off between accuracy and computation.

[1] E. Candes, L. Demanet, & L. Ying, 2007, Fast Computation of Fourier Integral Operators. SIAM Journal on Scientific Computing, Vol. 29, Iss. 6, pp. 2464-2493. 

[2] L. Demanet, M. Ferrara, N. Maxwell, J. Poulson, and L. Ying, 2012, A butterfly algorithm for synthetic aperture radar imaging. SIAM J. Imag. Sci., vol. 5, no. 1, pp. 203–243.

[3] S. I. Kelly and M. E. Davies, 2014, A fast decimation-in-image back-projection algorithm for SAR. 2014 IEEE Radar Conference, pp. 1046-1051.

[4] S. I. Kelly, M. E. Davies, J. S. Thompson, 2014, Parallel Processing of the Fast Decimation-in-image Back-projection Algorithm for SAR. 2014 Sensor Signal Procesing for Defence (SSPD),  pp. 1-5.

[5] V. Rokhlin and M. Tygert, 2008, A fast randomized algorithm for overdetermined linear least-squares regression. Proceedings of the National Academy of Sciences, 105(36), pp 13212–13217.

[6] P.-G. Martinsson and J. A. Tropp, 2020, Randomized numerical linear algebra: Foundations and algorithms. Acta Numerica, 29, 403–572.

[7] R. S. Srinivasa, M. A. Davenport and J. Romberg, 2019, Trading Beams for Bandwidth: Imaging with Randomized Beamforming. SIAM J. Imag. Sci., vol. 13, no. 1, pp. 317-350.

Industry Partner

This project is co-funded by Thales UK.

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. 

This project requires a student with a high level of mathematics or numeracy.

Further information on English language requirements for EU/Overseas applicants.

Full funding is available for this position.

Research Theme

Sensor Signal Processing

Aim

To develop machine learning algorithms for the robust detection and classification of natural and man-made objects in the ocean environment.

Objectives

1. Leverage unsupervised learning to understand the background ocean environment as a reference for detecting objects.
2. Develop a few-shot learning approach to classify the acoustic signals emitted by marine mammals and vessels travelling on the surface of the sea.
3. Explore strategies for improving the robustness of underwater acoustic models to distribution shifts.

Description

Passive sonar is a key sensing modality for achieving understanding of the ocean environment, to help protect naval platforms. This technology can be applied in a range of systems, from hull-mounted or towed hydrophone arrays to sonobuoys or uncrewed underwater vehicles. Sonar contact classification is a challenge, due to the need to detect, label, and track multiple targets. Much of this work is currently carried out manually by sonar operators. The growing risk from increasingly stealthy targets, complex environments, and a data deluge from more capable sensors with more channels, necessitates new automatic approaches to marine object detection and localization.

Recent developments in the fields of artificial intelligence (AI) and machine learning (ML) offer promise for improvements in the analysis of acoustic signals, in terms of speed, accuracy, and robustness. This project will investigate novel approaches to analyzing the data and provide an increased understanding of the maritime arena.

In later years of the PhD, QinetiQ would like to investigate the possibility of an internship for the student for up to 6 months.

QinetiQ has developed an algorithm for detecting the time stamps of sound cuts of short mammal calls in the master tapes of the Watkins Marine Mammal Sound Database. This enables the development and assessment of object detection algorithms as opposed to just classification of sound cuts, which is what the database is usually used for. The algorithm can be made available to the student to use and improve on.

QinetiQ has access to unclassified recordings of surface vessels. Subject to permission of the data owner, these could be made available to the student to support the research.

QinetiQ: https://www.qinetiq.com

Watkins: https://whoicf2.whoi.edu/science/B/whalesounds/index.cfm

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. 

Further information on English language requirements for EU/Overseas applicants.

Full funding is available for this position.

Research Theme

Sensor Signal Processing

Aim

To develop a framework for training adversarially robust machine learning models that addresses the uncertainty of adversaries’ capabilities and the difficulties of data collection in defence contexts.

Objectives

• Develop a learning algorithm for training adversarially robust Bayesian deep neural networks in the case where we have a pre-specified threat model.
• Show how to infer a threat model by analysing the distribution shift caused by deploying a system based on Bayesian neural networks.
• Use techniques from active learning and experimental design to minimise the amount of data required to accurately infer a threat model.

Description

Conventional machine learning methods are not designed to deal with the fog of war. Concerns from security and defence researchers have therefore lead to the development of adversarial machine learning methods. However, most prior work in this area focuses on very naïve threat models and make the assumption that the model trainer has perfect knowledge of the adversary’s capabilities. These naïve threat models often assume the adversary is only capable of manipulating some small number of pixels in an image, or that they can only add interference with small magnitude.

This project will develop general-purpose Bayesian machine learning algorithms for training robust models. The focus will be scenarios where the model trainer has minimal information about an adversary’s ability to manipulate sensor observations. We will show how analysing sequences or sets of distribution shifts induced by model deployments enables one to infer these unknown capabilities.  This will enable the deployment of adversarially robust models across a wide array of defence-relevant prediction problems (e.g., classification, detection, tracking) and data modalities (e.g., natural images, EO, SAR, acoustic, RF).

Minimum entry qualification - an Honours degree at 2:1 or above (or International equivalent) in a relevant science or engineering discipline, possibly supported by an MSc Degree. 

Prior knowledge of algorithmic game theory and probability theory would be advantageous.

Further information on English language requirements for EU/Overseas applicants.

Full funding is available for this position.