Fully-funded PhD Studentship – Deep learning for object detection, classification and analysis in non-photographic images using limited datasets

Fully-funded PhD Studentship

Deep learning for object detection, classification and analysis in non-photographic images using limited datasets

Robert Gordon University

Closing Date – Friday, 12 February 2021

Applications are sought for a fully-funded Research Studentship (PhD) to carry out research at Robert Gordon University, Aberdeen, United Kingdom, under the supervision of Dr Pamela Johnston and Professor Eyad Elyan.

Duration and Funding
The duration of project will be up to 36 months, commencing in Month and year. The studentship covers tuition fees (at Home or International (which includes EU students) level) and a tax-free stipend of £15,000 per annum.

Proposed Research
The proposed research is aimed at applying state-of-the-art computer vision techniques to a range of image data, including real-world datasets. Modern computer vision techniques involve deep learning. One of the major challenges in the application of deep learning is that many methods require large quantities of labelled data for supervised learning. Moreover, the best datasets are balanced and largely unbiassed. Datasets for deep learning are large, often including thousands of images and label data and the manual task of labelling images is an arduous one. Some public datasets exist, however these are often natural photographs which can be easily scraped from the internet and labelled by non-experts, e.g. ImageNet or Microsoft’s Coco. Real world applications can involve different types of images altogether, and may require expert labelling. Because of this, many real-world applications simply do not have large, available. Although dataset labelling can be somewhat accelerated with custom-built software tools, real progress can also be made by using the most recent machine learning techniques to expand existing small-scale datasets. Datasets can be enhanced by using GANs (Generative Adversarial Networks) to create realistic but synthetic examples of specific classes, already labelled and ready to use. Transfer learning can be applied to take machine learning knowledge from domains with large datasets and pre-trained models and apply it to real world situations with smaller but related datasets.

The supervisory team would like to examine the application of deep learning techniques to real-world projects with limited or imbalanced datasets. Applications might include: analysing complex engineering diagrams, individual cell segmentation and counting in cyanobacteria trichomes in micrographs, analysis of thermal images. Although seemingly distinct, these applications all involve real-world datasets with the common challenges of limited labelled data samples, non-photographic images and some degree of class imbalance. There is

the potential to solve some of these problems by using deep learning methods to generate synthetic, labelled data samples and evaluating their efficacy. Solving some of these challenges in one domain will help to advance research in other related domains.

This project will be focused mainly on:

  • Generating synthetic image data samples to tackle real-world dataset challenges and evaluating the efficacy of this
  • Identifying aspects of research applied to photographic images which can be transferred to non-photographic images
  • Exploring the use of unlabelled data in deep learning for computer vision

Key Skills

Applicants should have a very good BSc (Honours) (First or Upper Second class) degree or a Master degree (with Distinction or Merit) in Computing Science or related discipline.

Essential Knowledge and Experience:

  • Programming experience (preferably some in Python and some basic image processing techniques)
  • Strong communication skills (including written/spoken English)
  • Solid skills in maths/statistics
  • An ability to think independently and critically analyse different sources

Desirable requirements

  • Knowledge of deep learning techniques/packages (e.g. Keras, TensorFlow, PyTorch)
  • Key software development skills (e.g. version control, Bash, Docker)
  • Experience in the development, application and deployment of computer vision techniques

 Applicants should have good personal and communication skills, strong professionalism and integrity, and be capable of working on their own initiative.

Enquiries can be emailed to Kate Lines at soc-researchadmin@rgu.ac.uk and will be forwarded to Dr Pamela Johnston if technical in nature.

Applications

Applications should be emailed to Kate Lines at soc-researchadmin@rgu.ac.uk by midnight, Friday, 12 February 2021 (British time). The applications should consist of a covering letter or personal statement of interest, and a CV. Further information such as passport details or transcripts may be requested during the short-listing stage. Interviews (which may include a short practical test) are expected to take place week commencing 22 February 2021.