X-ray scanners are a commonly used tool to prevent contraband from entering secured facilities. Our research explores to what extent Convolutional Neural Networks can be used to automatically detect contraband in images generated from these scanners. We test the performance of Faster R-CNN, FCOS, EfficientDet and YOLOv5 on a dataset of X-ray images containing a test dummy equipped with different types of contraband, in combination with augmentation techniques like tiling, random affine transformations and pretraining. We find that Faster R-CNN (F1 = 0.593) and FCOS (F1 = 0.593) in combination with the usage of tiling and random transformations, and YOLOv5 (F1 = 0.570) in combination with pretraining are able to reliably detect some types of contraband, but fail to detect types of contraband which are mostly transparent to X-rays.
