Machine learning methods for discriminating natural targets in seabed imagery

Harrison, Richard, John Patrick (2012) Machine learning methods for discriminating natural targets in seabed imagery. Doctoral thesis, University of East Anglia.

[thumbnail of 2012HarrisonRJPPhD.pdf]
Preview
PDF
Download (8MB) | Preview

Abstract

The research in this thesis concerns feature-based machine learning processes and methods for discriminating qualitative natural targets in seabed imagery. The applications considered, typically involve time-consuming manual processing stages in an industrial setting. An aim of the research is to facilitate a means of assisting human analysts by expediting the tedious interpretative tasks, using machine methods. Some novel approaches are devised and investigated for solving the application problems.
These investigations are compartmentalised in four coherent case studies linked by common underlying technical themes and methods. The first study addresses pockmark discrimination in a digital bathymetry model. Manual identification and mapping of even a relatively small number of these landform objects is an expensive process. A novel, supervised machine learning approach to automating the task is presented. The process maps the boundaries of ≈ 2000 pockmarks in seconds - a task that would take days for a human analyst to complete. The second case study investigates different feature creation methods for automatically discriminating sidescan sonar image textures characteristic of Sabellaria spinulosa colonisation.
Results from a comparison of several textural feature creation methods on sonar waterfall imagery show that Gabor filter banks yield some of the best results. A further empirical investigation into the filter bank features created on sonar mosaic imagery leads to the identification of a useful configuration and filter parameter ranges for discriminating the target textures in the imagery. Feature saliency estimation is a vital stage in the machine process. Case study three concerns distance measures for the evaluation and ranking of features on sonar imagery. Two novel consensus methods for creating a more robust ranking are proposed. Experimental results show that the consensus methods can improve robustness over a range of feature parameterisations and various seabed texture
classification tasks. The final case study is more qualitative in nature and brings together a number of ideas, applied to the classification of target regions in real-world
sonar mosaic imagery.
A number of technical challenges arose and these were
surmounted by devising a novel, hybrid unsupervised method. This fully automated machine approach was compared with a supervised approach in an application to the problem of image-based sediment type discrimination. The hybrid unsupervised method produces a plausible class map in a few minutes of processing time. It is concluded that the versatile, novel process should be generalisable to the discrimination of other subjective natural targets in real-world seabed imagery, such as Sabellaria textures and pockmarks (with appropriate features and feature tuning.) Further, the full automation
of pockmark and Sabellaria discrimination is feasible within this framework.

Item Type: Thesis (Doctoral)
Faculty \ School: Faculty of Science > School of Computing Sciences
Depositing User: Users 2593 not found.
Date Deposited: 16 Jun 2014 14:41
Last Modified: 31 Jan 2016 01:38
URI: https://ueaeprints.uea.ac.uk/id/eprint/48800
DOI:

Downloads

Downloads per month over past year

Actions (login required)

View Item View Item