A Benchmark Dataset for Concealed Improvised Explosive Device Detection in X-ray Security Imaging.

Area of Science:

• Computer Science
• Security Engineering
• Image Analysis

Background:

• Automated threat detection in X-ray security screening is crucial for public safety.
• Existing Improvised Explosive Device (IED) datasets are limited, hindering machine learning advancements.
• IEDs pose a significant and persistent threat in high-security environments.

Purpose of the Study:

• To introduce IEDXray, an open-access dataset designed for automated IED detection.
• To provide a high-fidelity benchmark for evaluating machine learning models in security screening.
• To address the limitations of current IED datasets for developing advanced detection systems.

Main Methods:

• Constructed IEDXray with 17,360 X-ray images under a strategic concealment protocol.
• Included diverse IED types and varying clutter levels, simulating real-world baggage scenarios.
• Benchmarked state-of-the-art detection models (YOLOv10, Faster R-CNN, DETR, GroundingDINO) on the dataset.

Main Results:

• Established baseline performance metrics for leading object detection models on IED detection tasks.
• Demonstrated the dataset's capability to represent complex, real-world security screening challenges.
• Validated the utility of IEDXray for training and testing automated threat detection algorithms.

Conclusions:

• IEDXray serves as a valuable, open-access resource for the research community.
• The dataset facilitates the advancement of machine learning-based automated explosive detection.
• Improved detection capabilities can enhance the effectiveness of X-ray security screening systems.