Cookies

We use cookies to ensure that we give you the best experience on our website. You can change your cookie settings at any time. Otherwise, we'll assume you're OK to continue.

Computer Science

Prospective PhD Projects

There are various ways to come up with PhD proposals. Sometimes, academics announce a particular project. Sometimes, students write a draft proposal and refine it with a potential supervisor. Most often, potential students and potential supervisors tailor a given project sketch to the student's interests that fit into the research area of the respective supervisor.

In general, we welcome projects in the areas listed below. If you are interested in research overlapping with one of these fields, please contact the respective academic in Computer Science. We cannot accept research projects that have no overlap with any of our research fields. Please note, that in order to pursue any of these ideas as your research topic, you will need to secure suitable funding. For more information, please see the Research Degree Pages.



Algorithms and Complexity


Dr Maximilien Gadouleau

Problems in Discrete Mathematics for Scientific Computing (PhD)

Scientific computing solves continuous problems, such as differential equations, by discretising them. As such, problems somewhat similar to those in other branches of discrete mathematics, like graph theory coding theory, arise naturally. This project has two main aims. The first main aim is to identify such problems and to tackle them in a theoretical way. The second main aim is to validate our results and to evaluate the proposed solutions via the implementation of simulations using them.


Dr Maximilien Gadouleau

Algorithms for Memoryless Computation (PhD)

Memoryless computation is a new paradigm to compute functions. Most work has focused on evaluating how to compute a function in general. This project, however, aims at focusing on common classes of functions such as sorting, basic arithmetic etc. Designing memoryless programs for these functions is challenging and requires a lot of creativity. In the end, if efficient programs can be designed, this would provide a good case for the implementation of memoryless computation.


Dr Maximilien Gadouleau

Finite Dynamical Systems (PhD)

Many phenomena can be modelled as a Finite Dynamical System as follows. A set of agents interact with each other; each agent has a state and a function which updates this state; the function depends on the current state of some other agents. Gene regulatory networks, neural networks, or social networks can be modelled that way. The main aim of this project is to study the dynamical properties of those systems, i.e. how the states evolve over time. This study will use selected methods from a wide range of areas including graph theory, coding theory, dynamical systems, or algebra. Many dynamical properties are of interest, such as the number of fixed points, the period of the system, etc.


Dr Matthew Johnson

Complexity of Reconfiguration Problems (PhD or MSc)

The reconfiguration graph of a combinatorial decision problem contains as nodes all possible solutions to the problem, and two solutions are adjacent if their difference is minimal (for example, for the problem of colouring a graph, two solutions are adjacent if the colours differ on exactly one vertex). The proposed project might investigate finding algorithms (or hardness proofs) for the problem of deciding whether solutions are connected in reconfiguration graphs, but there are many other possible directions. There has been a lot of recent work in the area: to find out more consult this web survey


Professor Daniel Paulusma

Exploiting Graph Structure to Design Faster Algorithms (PhD or MSc)

Project Description: Many well-known graph-theoretic problems, such as graph coloring or finding a hamilton cycle, are computationally hard for general inputs but can be solved efficiently when we restrict the input to some special graph class. As an example, deciding whether a graph can be colored with at most four colors is an NP-complete problem. However, if the input graph is planar then the situation changes drastically, as the Four Color Theorem tells us that every planar graph is 4-colorable. The goal of this project is to design efficient algorithms for NP-complete problems restricted to special graph classes and to increase our understanding on such problems in general by researching and exploiting the underlying graph structure.


Dr Magnus Bordewich

Algorithms and Complexity in Phylogenetics (PhD or MSc)

Phylogenetics is the study of the relationships between species using DNA or protein sequence data. There are many computational and mathematical problems that arise in this field, such as: Given DNA data, what evolutionary tree best represents the true history of the species? Given an evolutionary tree, how efficiently can we determine which subset of species has the greatest biodiversity (as pictured)? Given an existing algorithm, what what is the relationship between the accuracy of the input data and the accuracy of the output? Such questions will be studied using discrete mathematics and algorithmic analysis. Suitable background for research in this field would be an undergraduate degree in computer science or mathematics. Examples of work conducted by a recent PhD student in this field are:

  • "On the fixed-parameter tractability of agreement-based phylogenetic distances.” Bordewich, M., Scornavacca, C., Tokac, N., and Weller, M. Journal of Mathematical Biology, 74:239-257 (2017).
  • "An algorithm for reconstructing ultrametric tree-child networks from inter-taxa distances.” Bordewich, M. and Tokac, N. Discrete Applied Mathematics, 213:47-59 (2016).

Links to these publications and further information can be found at: http://community.dur.ac.uk/m.j.r.bordewich/


Professor Iain Stewart

Interconnection Networks (PhD or MSc)

An interconnection network provides the communications fabric of a modern distributed-memory computer such as a many-core processor, a supercomputer, and a data centre. Whatever the context, interconnection network design is incredibly complex as there is a plethora of properties that we require our interconnection network to have, relating to scalability, flexibility, energy efficiency, latency, wiring complexity, throughput, load balancing, routing, fault tolerance, fault diagnosis, cost-to-build, reliability, and virtualization. However, some of these properties work against each other and so compromises have to be made, which makes the design of interconnection networks extremely challenging. The scale and cost of an interconnection network mean that we cannot simply build-and-test; hence, we use structural properties of the interconnection networks in tandem with simulation to evaluate our designs. As such, the design of interconnection networks is strongly linked with areas of discrete mathematics such as graph theory, coding theory, and even group theory. I am interested in all aspects of interconnection networks relating to their mathematics and theory. My research ranges from the combinatorial through to the simulation of large-scale data centre network designs (using a purpose-built simulator called INRFlow), and I am interested in developing PhD projects that range from the purely theoretical through to those that are more practical where the focus is more on simulation and validation using software.



Innovative Computing


Professor Toby Breckon

3D Computer Vision and Deep Learning within Autonomous Vehicles and Robotics (either PhD or MSc by Research)

The next generation of self-driving road vehicles and autonomous robotics will require accurate, real-time 3D scene information achievable on low-power consumption, high performance processors. Potential projects in this area can explore a range of options in this space including the design of low-power stereo vision approaches, deep learning for scene understanding (convolutional neural networks), enhancements to standard consumer level-depth sensing and real-time dense structure within low-power computational bounds. The applicant will participate in the ongoing programme of robotic sensing and autonomous vehicle research work concentrating on the real-time use of image-based visual sensing for navigation and environment understanding based on sensor test vehicles operated by Durham University. This may include aspects of real-time 3D scene reconstruction, semantic scene understanding and the challenges of all-weather operating conditions for intelligent robotics and driver assistance systems.

Related video content:

Keywords: Computer vision, image processing, robotic sensing, autonomous cars


Professor Toby Breckon

Autonomous Wide Area Search and Survey for Future Unmanned Aerial Vehicles (either PhD or MSc by Research)

Generalised wide are search and surveillance is a common tasking for the next generation of Unmanned Aerial Vehicles (UAV, “drone”) platforms. Whilst a range of techniques focus on topics such as target classification, recognition and tracking visual saliency is commonly overlooked within this context. By contrast, techniques such as visual saliency defines a measure or quality by which a given object or scene regions stands out relative their general context. Within a search and surveillance this can be used to automatically identify generic “objects of interest” without requiring a target-specific detection capability (e.g. people, vehicles), including those which are difficult to descriptively bound such as wreckage etc. Additionally this can be used to inform selective image transmission or encoding from UAV platforms in flight. Alternative approaches may explore applications of machine learning and “on-the-fly” 3D reconstruction within this problem domain. Current approaches do not consider aspects of temporal saliency, multi-modal saliency or 3D saliency over different sensing techniques or aspects of object shape giving ample scope for novel development. The applicant will participate in the ongoing programme of robotic sensing and computer vision research, underpinned by recent advances in machine learning, to consider such aspects within a wide-area UAV operation and survey applications.

Related video content:

Keywords: Computer vision, image processing, robotic sensing, UAV, aerial sensing


Professor Toby Breckon

Combining Multi-modal Signals and Images for Generalised Scene Understanding via Deep Machine Learning (either PhD or MSc by Research)

True integrated multi-modal sensing is one of the remaining challenges in generalised machine perception with applications ranging from multi-media web search, content-based media indexing and generalised scene understanding for tasks like robotic sensing. Despite an increasing array of low-costs sensors (e.g. video, audio, 3D stereo, RGB-D, infrared), all too often varying such sensory inputs are combined in a somewhat adjunct or simplistic manner. No generalised approach to the fusion or co-classification of multi-sensory inputs currently exists. Projects in this area aim to look at recent advances in efficient multi-modal data representation, deep machine learning and sensor fusion to tackle some aspects of this problem. The applicant will participate in the ongoing programme of multi-modal sensing and computer vision research spanning tasks such as complex environment understanding from mobile robots, autonomous surveillance and information retrieval.

Related video content:

Keywords: Computer vision, image processing, robotic sensing, signal processing


Professor Toby Breckon

Deep Learning for Next Generation Object Recognition – Applications within Airport and Border Security (either PhD or MSc by Research)

Future airport and border security screening systems will use increased levels of automation to improve performance, increase overall screening coverage and reduce security-related operating costs. “Within transport security, screening personnel are required to manually inspect thousands of baggage items for a range of contraband on a daily basis. In addition to this enormous workload, X-ray baggage imagery can be extremely challenging to interpret. Due to the nature of packed baggage, where objects are tightly packed, X-ray imagery generally contains a very high degree of clutter and inter-object occlusion. Consequently, objects are most often occluded or shown from unusual viewpoints. It has been shown that both human and computer detection rates are severely affected by complexity and clutter and therefore image interpretation in such environments is particularly challenging. Furthermore increasing global travel demands ever increasing turnover rates at security checkpoints allowing screening personnel only limited time to examine each baggage item" (Kundegorski/Breckon 2016).

This project aims to investigate recent developments in deep learning applied both to this specific object recognition application and more generally across a range of related object recognition challenges to investigate and advance the performance of current approaches. The applicant will participate in the ongoing research programme in this area, underpinned by recent advances in machine learning, and with direct access to X-ray security scanners in the research labs at Durham University.

Related video content:

Keywords: Computer vision, image processing, robotic sensing, signal processing, machine learning


Professor Toby Breckon

Deep Learning for Next Generation Automated Visual Surveillance – operating beyond the visible spectrum (either PhD or MSc by Research)

Infrared thermal imagery currently poses significant advantages for 24/7 day/night surveillance in terms of the visibility of human, animal and vehicle targets under all environmental conditions. However, a key limitation is the lack of colour detail suitable for many emerging approaches across the automated visual surveillance space – long term tracking, visual analytics, fine-grain classification, camera-to-camera tracking and alike.

Work on this theme aims to investigate the advancement and transferability of current state of the art techniques in this domain to address these issues across both infrared (thermal) and combined infra-red/colour imagery where available. Specifically, this theme aims to investigate recent developments in deep learning applied both to this specific visual surveillance application and more generally across a range of related target tracking and behaviour recognition challenges to investigate and advance the performance of current approaches. The applicant will participate in the ongoing research programme in this area, underpinned by recent advances in machine learning, and with direct access to on-site sensor hardware deployable on-site at Durham University and elsewhere.

Related video content:

Keywords: Computer vision, image processing, robotic sensing, signal processing, machine learning


Professor Gordon Love

Stereoscopic and Lightfield Displays, and their effect on Vision

How can we build a high fidelity display that replicates all the visual effects experienced by the eye in the real world?

This is a huge question encompassing many different factors. Much of the technological developments behind displays has concentrated on a relatively small number of (very important) factors such as resolution, dynamic range, and colour. We are interested in some of the other factors – especially involving the viewer and the eye which are particular important for stereoscopy. Stereoscopic (3D) displays have gone in and out of fashion over the years but they have seen a recent resurgence; especially for immersive technologies. Our research involves looking at the display cues which affect 3D vision in order to build better 3D and lightfield displays. Our projects are very interdisciplinary and involves Computer Science (Graphics), Physics (Optics and Imaging), Engineering (Display Technology) and Psychology (Vision Science). Current projects which we are working on involve Focus-correct stereoscopic displays, How can we correctly simulate blur (which is actually much more interesting and important than it sounds!) and How does the eye know how to accommodate (focus) correctly?


Dr Boguslaw Obara

BioImage Informatics for Spatio-Temporal Biological Networks (BIONET) (PhD)

The project aims to:

  • develop intensity-independent image analysis and processing solutions to extract and characterise the architecture of structural biological networks from 2D/3D/3D time-series images;
  • validate the proposed approaches using images of fungal, leaf vein and cytoskeletal networks with 10^6 of links across a range of physical scales;
  • build a unique benchmarking repository of complex biological networks with their topological characteristics.

The approaches developed here will enable robust extraction and quantitative characterisation of the architecture of 2D/3D/3D time-series biological networks. These quantitative measures will allow researchers to understand in which way topology and functions of the biological networks are related. This will then open new avenues, especially for researchers exploring the importance of fungal networks in causing diseases in crops, and of leaf veins and cytoskeletal networks in plant growth. Most importantly, adaptation of the proposed approaches need not be limited to biological images but can be applied to any images that contain curvilinear features. Specifically, the approach for a low-contrast feature extraction will be extremely beneficial to both the academic and industrial computing and bioimaging communities, as it will allow the confident use of low-contrast features in a wide range of different domains, such as biomedical imaging, robotics, astronomy, security and art, where image processing methods also play an essential role.


Dr Boguslaw Obara

Active Mesh-Based Segmentation and Objects Tracking (MESH) (PhD)

Deformable models are a group of algorithms used to segment images, particularly images with weak or missing boundaries. Such deformable models, also known as active models and snakes, are used heavily in biomedical imaging segmentation where low contrast,high noise or obscured edges make other segmentation techniques unsuitable. Deformable models come in several forms and have been used for multiple modalities and scenarios.

Active meshes are a subset of deformable models that commonly use a triangular-faced mesh to segment an object of interest from an image with weak or missing boundaries. The use of a mesh with distinct nodes, or vertices, and a known connectivity, through face or edge information, allows computationally simple and speedy calculations to be performed over the whole mesh. To find the segmentation shape the mesh is deformed. Active meshes tend to work on an iterative system with each iteration deforming the mesh to minimise energies or balance forces, eventually the mesh reaches an optimal state and the desired object is segmented. With many optimisation techniques there may be no definite and/or optimal termination; as such, many systems are provided with an explicit stopping criterion in order to terminate the process.

The project aims to:

  • develop active mesh-based image segmentation solutions to extract and characterise the complex objects from 2D/3D/3D biological images;
  • validate the proposed approaches using images of zebrafish lens cells.

Dr Boguslaw Obara

Multi-Scale Blob Detection in 2D/3D/4D Images (BLOB) (PhD)

In the area of image processing, blob detection refers to algorithms that are aimed at detecting points or regions in the image that differ in properties like brightness or color compared to the surrounding. There are several motivations for studying and developing blob detectors. One main reason is to provide complementary information about regions, which is not obtained from edge detectors or corner detectors. Blob detection is used to obtain regions of interest for further processing. These regions could signal the presence of objects or parts of objects in the image domain with application to object recognition and/or object tracking. Another common use of blob descriptors is as main primitives for texture analysis and texture recognition. Blob descriptors have also found increasingly popular use as interest points for wide baseline stereo matching and to signal the presence of informative image features for appearance-based object recognition based on local image statistics. There is also the related notion of ridge detection to signal the presence of elongated objects.

The project aims to:

  • develop multi-scale blob detector with automatic scale selection to recognize and track objects in 2D/3D/3D biological images;
  • validate the proposed approaches using images of zebrafish lens cells.

Dr Tobias Weinzierl

Data Aware High Performance and Scientific Computing

State-of-the-art algorithms in scientific computing often exhibit complex data access patterns. Examples are predictor-corrector schemes where the algorithm performs some massive calculations on chunks of small data and then corrects the solution by exchanging (few) information with other data chunks, or multilevel solvers where an algorithm successively coarsens the data representations (zoom out effect) and thus applies comparable operations on different data resolutions. Such algorithms have to be very carefully implemented if they shall run on big data sets and on massively parallel computers.

A typical PhD research project picks out one challenging algorithm, and studies how we can rearrange computations, decompose the problem on a supercomputer, perhaps introduce some optimistic calculations, hide data exchange, and so forth. Ideas developed then are prototypically implemented and tested before we try to formalise and to generalise them.

 
 

Contact Us

T:+44 (0) 191 334 1700

E: cs.office@durham.ac.uk