Real-time CNN-based object detection of prohibited items for X-ray security screening

Abstract

In homeland security, X-ray security scanners are essential for detecting contraband, including weapons, drugs, and hazardous materials. However, detection accuracy might be limited when items overlap or are concealed in complex baggage configurations. To address this issue, we developed a one-stage convolutional neural network (CNN) object detection model based on the You Only Look Once (YOLO) architecture, which employs a CSP-Darknet53 backbone for feature extraction. This model was designed for X-ray security scanners to enhance the detection of prohibited items in carry-on baggage, even under challenging conditions. The model was trained on a dataset comprising 6000 annotated X-ray images of suitcases, simulating real-world baggage scenarios with varying degrees of object overlap. Prohibited items were categorized into 22 classes based on the relevant International Civil Aviation Organization (ICAO) standards and Korean law, covering sharp objects (e.g., scissors, knives), weapons (e.g., guns, shurikens), and hazardous materials (e.g., liquids, flammables). To improve generalization, extensive image augmentation techniques were applied during training. Performance was evaluated using AP@50 (average precision at an IoU threshold of 0.5) and the F1-score, averaging 0.87 and 0.86 across all classes, respectively. Additionally, the average inference time per baggage item was 12.58 ms, making the system suitable for real-time airport security applications. These results demonstrate the potential of artificial intelligence (AI)-based security screening as a practical, high-speed detection solution for real-world airport security applications. This study addresses occlusion challenges in security screening by integrating multi-stage object overlap handling and dual-energy imaging. Compared to existing approaches, it could have higher applicability in complex baggage screening environments. While the model effectively detects structured prohibited items, further improvements are needed to identify non-rigid contraband, such as explosives and narcotics. Future research will explore dual-energy signal-based algorithms to enhance the detection capability of illicit substances.