Abstract
Surface determination process is critical step in metrological application with X-ray Computed Tomography(CT), that is performed post-reconstruction. Surface determination process approximates the location of surfaces through edge detection in X-ray CT images, generally attempted through ISO-50 method and local contrast search algorithm. This process would typically involve extensive computing processes, which can take hours to compute. This computing issue often cause a cycle time bottleneck in production setup, making XCT unsuitable as a large batch measurement tool. This thesis attempts to answer whether machine learning can be a feasible tool to speed up the current surface determination process while maintaining segmentation accuracy such that the output quality is usable for metrological operation. Throughan extensive literature search, it was determined that machine learning approaches have consistently outperformed rule-based approaches and have dominated state-of-the-art since 2015; however, it was determined that none of the approaches proposed in the literature is suitable for the surface determination process as it lacks the ability to segment a surface to a sub-pixel accuracy; trait that is critical for measurement applications. In this thesis, a novel lightweight deep learning architecture, with the capability to perform a sub-pixel segmentation of X-ray CT slices was proposed. The sub-pixel capability of the network was achieved through a convolutional neural network approach with multi-tasking framework. The architecture adapted a U-NET backbone, with multi-headed decoder. The job of the first head is to find the approximate location of the edge to a pixel level through semantic segmentation, and the second head offsets the prediction location based on a vector map created by regression. The model was trained on a
mixture of simulation, and real data. The inference result was compared against the current VGStudio’s advanced surface determination method as a reference surface. The result suggests that, against the reference surface, the inference result yields an average deviation of 3.25 ±1.9 (σ) pixels across all samples. The ISO-50 method yields an average of 13.4± 19.4 (σ) pixels. The architecture was also seen to be able to perform detection at a speed of 40 frames per second on a consumer GPU, completing full 2000 CT-stacks in less than two minutes. In comparison, VGStudio advanced surface determination method completed the computation in five to thirty minutes with 6 x NVIDIA1080-Ti GPUs, with the speed directly correlated to the complexity of the component and the noise that is present in the dataset. The result also shows significant improvement in data quality in the presence of severe X-ray scattering, beam hardening and cone beam artefacts. At the same time, it was observed that ISO-50 approaches faced difficulty in surface determination in the same dataset, resulting in random noisy points that lead to unusable surfaces for measurements; the novel approach showed that it is adaptive to identify the correct edges in such circumstances. The novel approach proposed in this thesis addresses the gap in the literature that prevents deep learning approach to be used in a metrological application for X-ray CT data due to the lack of sub-pixel edge detection properties. Moreover, the proposed model does not require any user input for reference, eliminating potential uncertainties caused user-dependency; an attribute that is common in heuristic approaches. The near real-time inference speed also reduces overall measurement time significantly compared to that of the traditional approach, paving the way for a full dimensional inspection in a largequantity production environment. This breakthrough marks a significant step towards a datadriven metrology approach for X-ray CT to be implemented with high reliability and computing efficiency.
| Date of Award | 28 Aug 2025 |
|---|---|
| Original language | English |
| Supervisor | Jane Jiang (Main Supervisor) & Shan Lou (Co-Supervisor) |