Threat Object-based anomaly detection in X-ray images using GAN-based ensembles.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Systems

Background:

• Detecting dangerous objects in luggage is critical for security at airports and other sensitive locations.
• Current methods rely on deep learning on X-ray images or human inspection, facing challenges with limited threat data.
• Anomaly detection using normal data samples offers a promising solution to the data scarcity problem in threat object detection.

Purpose of the Study:

• To develop a robust anomaly detection system for identifying dangerous objects in X-ray imagery.
• To address the challenge of limited high-quality threat image data in practical security scenarios.
• To enhance the performance of generative adversarial networks for computer vision-based anomaly detection.

Main Methods:

• Adopted and modified the Skip-GANomaly architecture with a UNet++ generator for improved performance.
• Developed an ensemble model combining Skip-GANomaly and the modified architecture for enhanced latent space exploration.
• Utilized Uniform Manifold Approximation and Projection (UMAP) for visualizing latent space and demonstrating model explainability.

Main Results:

• The modified Skip-GANomaly achieved an Area Under the Curve (AUC) of 94.94% on the Compass-XP dataset.
• The proposed ensemble model achieved a superior AUC of 96.8% on the Compass-XP dataset.
• The ensemble model demonstrated better feature learning for anomaly separation compared to individual architectures, validated on the SIXray dataset.

Conclusions:

• The developed ensemble architecture provides state-of-the-art results for threat object detection in X-ray images.
• The models can effectively detect threat objects without prior training on images containing such objects.
• The approach offers a promising solution for security screening by overcoming data scarcity challenges in anomaly detection.