3D localisation of sparse internal features with machine learning based stereo X-ray tomography
X-ray computed tomography (XCT) is a well-established volumetric imaging technique that is widely used in scientific research and industrial inspection. However, the slow acquisition process in modern XCT systems remains a fundamental limitation. This is mainly caused by the relatively long exposure time required for each individual projection and the large number of projections needed over different viewing angles to achieve full tomographic reconstruction. Motivated by early stereoscopic X-ray imaging, concepts from visible-light stereo vision, and recent advances in deep learning, this thesis investigates a learning-based stereo X-ray tomography framework for three-dimensional localisation of sparse internal features using only two projection images. Instead of performing full volumetric reconstruction, the proposed framework focuses on the direct 3D localisation of point- and line-based features, thereby reducing data acquisition requirements for this constrained localisation task. Building upon this framework, the method is further extended to more complex geometric features, including edges and corners, under controlled component-specific imaging conditions. To explore the potential of stereo X-ray tomography for time-resolved imaging, the framework is also applied to simulated deformable-object tracking, where feature positions are estimated sequentially from consecutive stereo image pairs. The results presented in this thesis demonstrate that stereo X-ray tomography is a promising proof-of-concept approach for sparse feature localisation and tracking, particularly in settings where rapid imaging and reduced acquisition time are important. Further validation across independent samples, imaging conditions, and real dynamic experiments is required before broader practical deployment.