X-ray Images Augmented with Simulated Virtual Flaws for Deep Learning Based Defect Detections

Abstract This article analyzes the performance of object detection algorithms in detecting flaws in welds inspected by X-ray testing. We propose a learning framework that relies on both experimental acquisitions and simulations to build a training set of images with virtual flaws. This set is then used to fit object detection algorithms aiming at detecting flaws in welds. After training, we analyzed the machine learning performance against real flaws in welds using experimental X-ray testing images that were not included in the training process.