Domain-inspired image processing and computer vision to support deep-sea benthic ecology

Game, Chloe (2022) Domain-inspired image processing and computer vision to support deep-sea benthic ecology. Doctoral thesis, University of East Anglia.

[thumbnail of CG 231113 Final PhD Thesis.pdf]
Preview
PDF
Download (35MB) | Preview

Abstract

Optical imagery is a necessary methodological tool for ecological research within marine environments, particularly in deeper waters. For benthic (seafloor) surveys, interpretation of image data is crucial to creating high-resolution maps of seabed habitats. This is fundamental to marine spatial planning and mitigating long-term damage of anthropogenic stressors such as growing resource demand, climate change and pollution. However there are numerous, and significant, issues in extracting a reliable ground-truth from imagery to support this process.

Analysis of benthic images is difficult, due in part to the extreme variation and inconsistency in image quality - caused by complex interactions between light and water. It is also time-consuming. This thesis is dedicated to providing solutions to manage these challenges, from a strong perspective of the end-user. Specifically, we aim to improve the annotation of benthic habitats from imagery in terms of quality, consistency and efficiency. Throughout, we consider the purpose the imagery serves and work closely with end-users to best optimize our solutions.

First, and for the majority of this thesis, we investigate image processing techniques to improve the appearance of image features important for habitat classification. We find that tone mapping is an effective and simple (and thus accessible) method through which to improve image quality for interpretation. We describe beneficial (expert-informed) properties for brightness distributions in underwater images and introduce a novel tone-mapping algorithm, Weibull Tone Mapping (WTM), to enhance benthic images. WTM theory operates within general constraints that model image requirements (properties) specified by image analysts, yet possesses a suitable degree of flexibility and customisation. As a tool, WTM provides analysts with a fast and ‘user-friendly’ method to improve benthic habitat classification.

Second, we consider computer vision methods that could automatically identify benthic habitats in imagery, relieving the analysis bottleneck. We find that baseline transfer learning of machine learning models, with limited optimization, will better facilitate adoption by novice users, yet still provides a powerful means to swiftly extract and assess benthic data.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Chris White
Date Deposited: 12 Dec 2023 12:01
Last Modified: 12 Dec 2023 12:01
URI: https://ueaeprints.uea.ac.uk/id/eprint/93971
DOI:

Actions (login required)

View Item View Item