Thesis

Signal and image processing for enhanced long range sensing : improving long range sensing to counter developing threats

Downloadable Content

Download PDF
Creator
Rights statement
Awarding institution
  • University of Strathclyde
Date of award
  • 2026
Thesis identifier
  • T17661
Person Identifier (Local)
  • 201790727
Qualification Level
Qualification Name
Department, School or Faculty
Abstract
  • Methods for rapidly identifying and discriminating between objects of interest and their background is key in many applications. With the increasing prevalence of hyperspectral cameras and the large datasets they produce, the need for efficient and interpretable tools for capturing, processing and analysing such data is imperative. The work presented in this thesis focuses on several techniques for improving upon existing methods of object and target detection in electro-optical imaging modalities. This thesis is split into three distinct parts linked by the common theme of improving target detection in hyperspectral imagery. The first part investigates methods of extending the morphological Hit-or-Miss Transform for use in multivariate images as an efficient and explainable object detection algorithm. The second investigates methods of improving target detection in aerial hyperspectral imaging applications through the use of joint spatial and spectral dimensionality reduction. The final section investigates the use of simulated compressive sensing hardware for capturing a compressed representation of a scene and performing target detection and localisation without the need for any prior reconstruction. The first contribution chapter focuses on extending a morphological approach to object detection, the Hit-or-Miss Transform. Traditionally constrained to single channel imagery, the proposed extension leverages the colour or spectral information afforded by multivariate data to detect objects based on their similarity to a reference in this high dimensionality space, as well as their size and shape. By incorporating the notion of percentage occupancy, the proposed method is made robust to noise and occlusion. Using various synthetic and natural images, the performance of the proposed Multi-Dimensional Percentage Occupancy Hit-or-Miss Transform is presented and compared with similar techniques showing that it can be more reliably applied to target detection tasks in multivariate data. The second contribution chapter details a target detection pipeline using optimal spatial and spectral dimensionality reduction techniques. Target detection and classification is an important application of hyperspectral imaging in remote sensing as the increased spectral resolution allows for potentially greater distinction between target and non-target pixels in an imaged scene. Both the high redundancy, and sparsity of targets, in hyperspectral image data can be exploited. By eliminating pixels with known, non-target, characteristics prior to spectral dimensionality reduction, not only can the size of the data be decreased, but the representation of targets within the reduced domain is improved. The proposed approach, Joint Spatio-Spectral Dimensionality Reduced Target Detection, achieves > 95% compression rates whilst preserving target detection performance in aerial hyperspectral images. The third, and final, contribution chapter presents a method of performing target detection on data collected using Compressive Spectral Imaging without the need for prior reconstruction. This has the potential to save on the often costly and time intensive capture, storage, or transmission of highly redundant data, which is then largely discarded. Coded Aperture Snapshot Spectral Imaging captures compressed measurements of a scene via a known optical architecture based on the design of a coded aperture, as a result the behaviour of some desired target spectrum through the same architecture can be predicted. An approach for exploiting this, Coded Aperture Snapshot Spectral Imaging for Target Detection is presented as a baseline for performing target detection from compressed measurements and obtains promising results on synthetic and natural multispectral imagery. The work presented in this thesis aims to improve upon target detection in hyperspectral imagery by exploiting the inherent sparsity in the data. Each contribution investigates one of many approaches to improving target detection, namely; efficient detection algorithms, optimal data compression, and imaging hardware. Demonstrating that this multifaceted problem, requires a multifaceted solution.
Advisor / supervisor
  • Murray, Paul
Resource Type
DOI
Funder

Relations

Items

Thumbnail Title Date Uploaded Visibility Actions
file details PDF of thesis T17661 2026-03-10 Public Download