딥러닝 기반 X-선 보안검색 시스템 개발 = Development of an X-ray Security Screening System with Deep Learning

다국어 초록 (Multilingual Abstract)

X-ray security screening is essential for preventing contraband, including drugs, weapons, and explosives, from being smuggled into major facilities like airports and customs. However, conventional X-ray security screening is often ineffective at detecting organic contraband, such as drugs or explosives, because these items lack a distinctive structure. Furthermore, operators typically have only a few seconds to interpret images, and accuracy depends heavily on their experience. This study aims to address these limitations by developing a real-time, automated interpretation framework based on deep learning.
First, the radiation generator and dual-energy detector were optimized using the figure of merit (FOM). An 80–160 kVp X-ray spectrum was then generated using MCNP6, after which the detector was evaluated using Geant4-based simulations. The System's performance was then validated using the ASTM F792-HP test kit (e.g., Wire Display and Spatial Resolution). This demonstrated an improvement in Wire Display of more than 30% compared to a commercial system. The discrepancy between the experimental measurements and the simulations remained within 5%, confirming the reliability of the modeled system.
A model for detecting prohibited items based on deep learning and using the YOLO architecture was subsequently developed. According to ICAO regulations and Korean aviation security standards, 22 classes of prohibited objects were defined, including weapons and liquids. The model was trained using 6,000 baggage images with extensive data augmentation. The model achieved an overall detection accuracy of 87% with an average inference time of 12.58 ms per bag. While the model is effective at identifying contraband of a specific shape, it struggles to detect irregular organic materials such as explosives and drugs. To address this issue, a CNN-based material classification algorithm was developed to directly differentiate between six material categories—background, organic, inorganic, metal, drug, and explosive—at the patch level. Unlike conventional systems, which rely on R-value-based pseudo-coloring and cannot reliably separate drugs or explosives from everyday organic items, the proposed model achieved an overall accuracy of 97.9%. Detection performance for the drug and explosive classes was 95% and 99%, respectively. Further testing using a multi-material phantom showed that the algorithm retained spatial accuracy and outperformed traditional formula-based discrimination methods.
These results demonstrate that deep learning-based object detection and material discrimination can identify illegal items in real-time, even in complex environments such as airports and customs facilities. These techniques can significantly enhance the reliability and efficiency of security operations by addressing the limitations of traditional X-ray screening methods.

목차 (Table of Contents)

• I. INTRODUCTION 1
• II. DUAL-ENERGY X-RAY SCREENING SYSTEM 3
• 2.1 Baggage Screening in Homeland Security 3
• 2.2 Dual-Energy Transmission X-ray Imaging 5
• 2.3 Performance Evaluation for Security Screening Systems 9
• 2.3.1 Evaluation with Standard Test Kits 9
• 2.3.2 National Standards for Evaluation 13
• 2.4 Limitations of Conventional Security Screening 15
• Ⅲ. DEVELOPMENT OF AN X-RAY SECURITY SCREENING SYSTEM 17
• 3.1 Design Optimization of the Security Scanner 17
• 3.1.1 Design Optimization Procedure 17
• 3.1.2 Radiation Generator 21
• 3.1.3 Dual-Energy X-ray Detector 24
• 3.2 Preliminary Performance Prediction through Simulation 29
• 3.2.1 Preliminary Prediction using Standard Parameters 29
• 3.2.2 Material Classification with a Simplified Phantom 32
• 3.3 Experimental Validation of the Optimized Design Model 35
• 3.3.1 Experimental Setup and Simulation Model 35
• 3.3.2 Experiments using a Radioactive Isotope 38
• 3.3.3 Experiments using a Radiation Generator 42
• 3.4 System Deployment and Field Testing 52
• 3.4.1 Test Bed for Trial Operation 52
• 3.4.2 On-site Evaluation at a Customs Facility 55
• 3.5 Performance Evaluation of the Developed System 58
• 3.5.1 Experimental Setup and Simulation Model 58
• 3.5.2 Wire Display Performance 61
• 3.5.3 Spatial Resolution Performance 66
• 3.5.4 Useful Penetration Performance 70
• Ⅳ. DEVELOPMENT OF DEEP LEARNING-BASED THREAT DETECTION ALGORITHMS 78
• 4.1 X-ray Security Image Interpretation Models 78
• 4.2 Object Detection Model for Prohibited Items in Aircraft 80
• 4.2.1 Overview of Object Detection Algorithms 80
• 4.2.2 Prohibited Items in Carry-On Baggage 82
• 4.2.3 YOLO-based Pre-Trained Models 85
• 4.2.4 Dataset of Carry-On Baggage 87
• 4.2.5 Fine-Tuning for Prohibited Item Detection 91
• 4.2.6 Performance Evaluation of Object Detection 97
• 4.2.6.1 Overlap-Stage Dataset 97
• 4.2.6.2 Realistic Scenarios for Suitcase Scanning 101
• 4.3 CNN-based Model for Illicit Material Signal Detection 106
• 4.3.1 CNN-based Material Classification Framework 106
• 4.3.2 Stride-based Image Reconstruction Algorithm 112
• 4.3.3 Dataset of Illicit Material Signals 117
• 4.3.4 Model Development and Optimization 120
• 4.3.5 Patch Size Optimization 122
• 4.3.6 Performance Evaluation of Material Classification 126
• 4.3.6.1 Evaluation using Reconstruction Phantoms 126
• 4.3.6.2 Evaluation using Multi-Material Test Phantoms 128
• Ⅴ. CONCLUSIONS AND FUTURE WORK 131
• 5.1 Summary and Conclusions 131
• 5.2 Suggestions for Future Work 133
• References 134
• Abstract (In Korean) 144
• Acknowledgements (In Korean) 146