Agenda
July
-
14:0018:00
In the case of additive manufacturing, X-ray tomography inspection technology appears to be a highly relevant approach for addressing the inspection needs of mechanical parts, given their geometric complexity (internal functional shapes), material structure, and the types of defects inherent to these technologies. Analyzing tomography data can be time-consuming and costly for large batches of parts due to the significant allocation of human resources, including for tasks with low added value that are sometimes repetitive. Automation should also allow for the focused deployment of skilled operators when dealing with complex findings that require classification—that is, on highly critical operations with high added value. This thesis proposes a hybrid data augmentation methodology that operates directly in the projection domain. The approach exploits the mathematical properties of the Radon transform to extract, model, and inject defects in a controlled manner into a target volume, without requiring prior volumetric reconstruction. It is based on three complementary modules: iterative extraction of sinusoidal defect signatures from sinograms, realistic 3D surface reconstruction of defects, and a synthesis methodology for target industrial parts. The databases are automatically generated from synthetic sinograms of realistic defects. Tailored for deep learning, they have been used to train and evaluate segmentation models such as UNet and YOLO, demonstrating the effectiveness of sinogram-based volumetric augmentation in improving segmentation for both simple and complex geometries.
Amphi LaBRI