In our 2D X-ray non-destructive testing (NDT) pipeline, we use two inspection
strategies: (a) reference-less inspection, which works well on simple parts; and
(b) reference-based inspection, which is used for complex, large parts where
reference-less methods result in false positives and missed defects. Reference-
based detection fundamentally relies on a high-quality ’golden’ image that must
be precisely aligned with the test image for accurate defect detection. However,
current moment-based registration algorithms perform poorly when confronted
with practical imaging variations, including translation, rotation, and slight non-
rigid deformations. Slight changes in perspective (common in X-ray setups due
to varying source-detector distances) are not handled well, resulting in residual
misalignment. These registration failures directly result in critical defects being
missed or false positives being detected.
This thesis will identify and evaluate registration approaches that can handle
rigid transformations and slight scale differences while preserving small defect
artefacts.