Data efficient deep learning for automated visual environment monitoring

French, Geoffrey (2023) Data efficient deep learning for automated visual environment monitoring. Doctoral thesis, University of East Anglia.

[thumbnail of GF 240501 Final PhD thesis.pdf]
Download (28MB) | Preview


Deep neural networks have been used to establish a number of state of the art results in computer vision over the last several years. Their use has resulted in a step-change in performance in a variety of computer vision tasks, often raising accuracy over the threshold at which automated analysis becomes practically useful. This performance however comes at a cost requiring large labelled datasets for training.

Computer vision training sets typically consist of input images with corresponding ground truth annotations. Acquiring or producing them can be challenging. In some domains – e.g. medical imagery – the major difficulties are associated with acquiring input images due to the necessity of expensive equipment or complications due to privacy concerns. In many domains – particularly those involving photographic imagery – input images are readily available, while the manual process involved in producing the ground truth annotations acts as a bottleneck as it can be a laborious and expensive task.

In this thesis we will discuss practical computer vision problems where the cost of producing ground truth annotations poses a significant limitation. We explore and discuss a number of approaches aimed at reducing the quantity of ground truth labels required or reducing the effort required to produce them.

This work is motivated by practical problems related to environmental monitoring that will be discussed, along with the solutions that we were able to apply.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Nicola Veasy
Date Deposited: 26 Jun 2024 10:09
Last Modified: 26 Jun 2024 10:09


Downloads per month over past year

Actions (login required)

View Item View Item