Thesis
New deep learning techniques for image enhancement and classification
- Creator
- Rights statement
- Awarding institution
- University of Strathclyde
- Date of award
- 2025
- Thesis identifier
- T17496
- Person Identifier (Local)
- 202064128
- Qualification Level
- Qualification Name
- Department, School or Faculty
- Abstract
- Deep learning has revolutionised the field of image processing, enabling significant advancements in tasks such as classification, object detection, and image enhancement. However, several critical challenges persist, hindering its broader applicability and efficiency. This thesis investigates solutions to key issues, including the generalisation of models to new classes, the scalability of deep learning systems constrained by their substantial size, and the limitations of supervised learning in acquiring labelled data at scale. Additionally, it explores innovative approaches to improve image enhancement, with a focus on reconstruction fidelity and computational efficiency. This research contributes novel model architectures, training techniques and insights to the development of robust, efficient, and versatile deep learning frameworks for image processing. One such field where the generalisation of models is critical is the dairy industry. Automated identification of individual cattle is a valuable tool for modern dairy farming, enabling increased operational scale and the potential for advanced health and welfare monitoring systems. Existing identification methods, such as Radio Frequency Identification tags, achieve only around 90% accuracy, are prone to detachment, and require specific scanning locations. Recently, deep learning-based identification systems have gained attention for their ability to overcome these limitations. This thesis explores deep learning techniques for training cattle identification models using data acquired in a controlled milking parlour environment, aiming to enhance existing Radio Frequency Identification systems. Through similarity learning, models are trained to produce embeddings that enable identification of cows not present in the training set. Novel new class analysis is conducted to evaluate model performance in realistic scenarios where herd composition changes over time, demonstrating the generalisation capacity of this technique. Furthermore, cattle identification models trained on controlled milking parlour data excel in similar domains but struggle to generalise to free-moving barn environments, where labelling data at scale is impractical. To address this, novel self-supervised learning techniques are proposed in this thesis to facilitate domain adaptation. These techniques leverage detection and tracking models to generate weak labels from unlabelled barn data, which are then utilised in triplet loss functions during training, achieving significant performance improvements over existing self-supervised approaches. Another field which has seen massive advancements due to breakthroughs in deep learning is Hyperspectral imaging. Hyperspectral imaging is a valuable tool in remote sensing applications as its spectral properties offer rich insights into the materials present within each captured hyperspectral image. However, this spectral detail typically comes at the cost of reduced spatial resolution. To address this, Super-Resolution techniques are often used to recover lost spatial detail and improve the overall quality of hyperspectral images. Despite their potential, several challenges persist in this domain, including issues related to data quality caused by sensor noise and the spectral response of sensors. However, the most critical challenge specific to Super-Resolution, is the lack of paired high- and low-resolution training data. As a result, existing methods often rely on artificially generating low-resolution image pairs, leading to suboptimal performance in real-world scenarios. To address these limitations, this thesis introduces several key contributions to the field of Hyperspectral Image Super-Resolution, including a novel paired high- and low-resolution dataset, novel preprocessing techniques, and a novel analysis of models trained using synthetic downsampling methods and evaluated on the proposed datasets. The power of deep learning models comes from their ability to learn highly complex non-linear functions. However, the non-linear components of a deep neural network come from the activation functions used between layers, meaning that to achieve the necessary non-linear complexity, networks have to be sufficiently deep, resulting in large computational demands. Self-Organised Operational Neural Networks (Self-ONNs) have recently been proposed, which tackle this issue by making the linear filters of traditional Convolutional Neural Networks (CNNs) learnable non-linear functions, meaning that the same theoretical non-linear complexity can be achieved in a much shallower network. To address computational challenges, novel Self-ONN architectures are proposed. Architectures are proposed for both cattle identification tasks as well as Hyperspectral Image Super-Resolution to demonstrate both the power and versatility of such models. The results presented in this thesis show that more parameter-efficient Self-ONN models can achieve performance on par with larger CNN models and in certain cases, even outperform them. This thesis presents a comprehensive exploration of deep learning methodologies tailored for practical applications in image processing, offering contributions that span cattle identification, Hyperspectral Image Super-Resolution, and self-supervised learning for domain adaptation, paving the way for more robust and scalable solutions.
- Advisor / supervisor
- Murray, Paul
- Marshall, Stephen
- Resource Type
- DOI
Relations
Items
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
PDF of thesis T17496 | 2025-11-06 | Public | Download |