Automated Damage and Defect Detection with Low-Cost X-ray Radiography Using Data-Driven Predictor Models and Data Augmentation by X-ray Simulation

The detection of hidden defects in materials using X-ray images is still a challenge. Often, a lot of defects are not directly visible in visual inspection. In this work, a data-driven feature marking model is introduced to perform semantic pixel annotation. Input data are delivered by a standard industrial X-ray instrument and a low-cost self-constructed portable X-ray instrument, which is introduced in detail in this work, too. The technical details of the X-ray instrument are relevant since the quality of the feature detector is compared with respect to noise, contrast, and signal quality. Finally, a simulation of X-ray images is used to provide a ground truth data set for the training of the feature detector. It is shown that this approach is suitable for detecting hidden pores in high-pressure die-casted aluminum plates.