Micro-XCT Imaging and Synthetic Data Generation for Deep Learning-Based Computer Vision of Microstructural Defects

e-Journal of Nondestructive Testing, vol. 30

Abstract

This study introduces a pipeline that integrates high-resolution micro-X-ray computed tomography (micro-XCT) with synthetic image generation to enable deep learning-based computer vision of microstructural defects. Ammonite fossils were chosen as a case study to highlight challenges in conventional porosity analysis, which often suffers from thresholding artifacts. To address these limitations, we developed a modular framework that procedurally generates annotated synthetic images by combining realistic textures with geometrically varied foreground defects using Perlin noise and affine transformations. The resulting datasets were used to train a Mask R-CNN segmentation model in the OpenMMLab MMDetection framework. The trained model achieved strong mean average precision across IoU thresholds and consistently high recall, particularly for larger defects, exceeding the performance of standard image analysis workflows. The method generalized well to real XCT data, enabling accurate pore identification in complex fossil structures. This approach demonstrates a scalable, domain-adaptive solution for non-destructive inspection, with positive implications for improved quality control and microstructural analysis in materials science.

Authors 3

1. Ravshan Yormatov Aachen Access e.V

   Access

   Affiliation as printed

   Access e.V. , Intzestraße 5 , 52072 Aachen , Germany;

2. Angelos Theofilatos Aachen Access e.V

   Access

   Affiliation as printed

   Access e.V. , Intzestraße 5 , 52072 Aachen , Germany;

3. Alexandre Viardin Aachen Access e.V

   Access

   Affiliation as printed

   Access e.V. , Intzestraße 5 , 52072 Aachen , Germany;