A Physics-Driven X-Ray Image Data Augmentation Method for Automated Threat Detection in Nuclear Facility Entrancement

Abstract

Due to the perspective ability of the internal structure of objects, the X-ray security screening system has been used to ensure nuclear security in nuclear facility entrancement. To improve detection efficiency and accuracy, recent work posed a particular interest in automated threat detection algorithms based on deep learning. A significant obstacle to developing high performance of such algorithms is the difficulty of obtaining large labeled training datasets. X-ray image data augmentation strategies based on experimental images have been proposed by other investigators. This paper proposed a physics-driven X-ray image data augmentation method for automated threat detection. Using the 3D modeling software, we can construct threat models and project them into threat images without experiment. According to Lambert-Beer law, the simulation image datasets can be made by exponentially overlaying the threat images on the background images. Considering the energy spectrum and scattering effect, we further process the phantoms and projection images accordingly to be more authentic. We train YOLOv5 architecture on the simulation dataset and test the algorithm on the experimental images. The results show that our approach achieves good performances in automated threat detection with an average recognition accuracy of over 90% and a mAP@0.5 of 82.9%, which is an effective method to increase the X-ray image dataset in both quantity and diversity.