Deep Learning-based Artificial Neural Network Object Detection Methods for Inspection of Hazardous Items

To detect hazardous materials within cargo passing through ports and airports, X-ray images are captured and inspected by human screeners. However, to ensure detection accuracy and consistency, various object detection algorithms using deep learning neural networks are being applied in the video search for pre-screening of dangerous items. In this study, various detector algorithms, such as Faster R-CNN, PAA, and D2Det, with the Vision Transformer (ViT) backbone, were applied for detecting hazardous objects within X-ray baggage images. Additionally, several training techniques, including self-supervised learning methods like Soft Teacher, MAE, and SimMIM, as well as transfer learning from pretrained models using techniques like DeiT and BEiT, were analyzed to improve the detection performance. When applied for object detection using the SIXray dataset, the D2Det detector with multiway Swin Transformer backbone exhibited a detection performance with mean average precision (mAP0.5) of 86.1%, while the Faster R-CNN detector with ViT backbone pretrained with BEiT exhibited a detection performance with mAP0.5 of 85.5%.