Automatic threat detection in X-ray images using deep learning

Automated object detection in X-ray baggage screening is critical for maintaining security and operational efficiency in high-throughput environments such as airports. Traditional methods often struggle with the unique challenges of X-ray imagery, including overlapping objects and low contrast. Recent developments in deep learning, and the YOLO (You Only Look Once) architecture specifically, have been especially promising for real-time object detection. This thesis benchmarks the performance of the newest YOLO variants— YOLOv8, YOLOv9, and YOLOv10—on three X-ray baggage datasets that are widely used—CLCXray, PIDXray, and SIXray. The comparison is done based on important parameters such as detection accuracy, inference speed, and computational cost to evaluate their applicability in real-time applications. Comprehensive experiments are implemented to evaluate their performance in object detection in dense and complicated scenes with an emphasis on the trade-off between detection accuracy and processing time. Results indicate that YOLOv10 performs best overall with higher accuracy and quicker inference at low computational complexity. YOLOv8 and YOLOv9 also show competitive performance with benefits under certain circumstances. Results indicate the efficacy of recent YOLO models in meeting the requirements of real-world X-ray baggage screening systems and their readiness for deployment in operational security settings. This research provides a detailed analysis of cutting-edge object detection models, contributing immensely to the knowledge of their application in real-world scenarios and paving the way for the creation of more sophisticated automated security systems.