Abstract
Industrial X-ray Computed Tomography (XCT) represents a pivotal non-destructivetesting methodology for quality control across a range of industrial sectors. The accurate
segmentation of XCT data facilitates the identification of defects and the characterisation of materials. This thesis presents the results of an investigation into the application
of the Segment Anything Model (SAM) in this context.
SAM, a state-of-the-art machine learning model developed by Meta AI, employs
an unsupervised approach that automates segmentation without manual annotations
and combines deep Convolutional Neural Network (CNN) and generative adversarial
networks. SAM was evaluated on diverse industrial XCT data, demonstrating its competitive segmentation accuracy and efficiency. However, limitations were also identified,
including its instance segmentation and training on street images instead of XCT scans.
SAM’s unsupervised nature and adaptability offer cost-effective and reliable quality
assurance solutions in industrial settings, enhancing product quality and inspection
processes.
Theoretical frameworks from machine learning and computer vision are synthesized
with empirical findings to assess SAM’s efficacy and potential for real-world applications. Moreover, practical insights gleaned from experimentation shed light on SAM’s
performance across diverse industrial datasets, highlighting both its strengths and areas
for improvement. The findings of this research provide a framework for future research,
emphasising interdisciplinary collaboration, iterative refinement, and a commitment to
innovation in reshaping the landscape of industrial inspection and quality assurance.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Ulrich Bodenhofer (Supervisor) & Patrick Weinberger (Supervisor) |