On the joint use of simulations and experimental RX images for generating very large datasets with application to object detection tasks

Machine Learning (ML) and Deep Learning (DL) algorithms are increasingly used in industrial domains to enhance accuracy and efficiency in error-prone tasks. Their performance hinges on data quantity and quality. In Nondestructive Testing and Evaluation (NdT&E), data scarcity, often due to confidentiality and rare events, poses a challenge. This is particularly true for Radiographic Testing (RT) data. To address this, we propose a systematic approach to combine simulated data with experimental radiographic (RX) images, creating large datasets with common defects (e.g., porosity, flaws) found in nuclear welding inspections. This process, called in this work Virtual Flaws (VFs) insertion, enables training ML algorithms without manual labelling or extensive experimental datasets. We evaluate the effectiveness of VFs-based datasets in object detection tasks using state-of-the-art algorithms.