Synthetic X-ray Projections Generation from CT Data for Deep Learning Applications

X-ray digital radiography is a widely used non-destructive testing technique, valued for its speed, efficiency, and ability to visualize internal defects. Nowadays, there is a growing reliance on X-ray inspections in high-throughput or in-line quality control environments, motivating the adoption of neural networks for automatic defect recognition. However, the performance of such networks is often limited by the scarcity of high-quality annotated data and the complex, variable internal geometries of real-world objects. Generating realistic and diverse simulated training data is, therefore, essential for developing robust automated inspection systems. We present a method for augmenting training datasets using information derived from CT measurements In this approach, the natural morphology of real manufacturing defects, as extracted from CT data, is preserved. The segmented defects are converted into 3D STL models, randomly repositioned and resized, and subsequently used to simulate diverse projection data under controlled conditions. This enables the creation of extensive, physically meaningful training datasets that capture a wide range of defect variations while faithfully preserving their true characteristics.