Meta-TIP: An Unsupervised End-to-End Fusion Network for Multi-Dataset Style-Adaptive Threat Image Projection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• Threat Image Projection (TIP) is crucial for training security personnel and AI systems for X-ray baggage screening.
• Existing TIP methods, principle-based and GAN-based, have limitations in flexibility, style consistency, training stability, and interpretability.

Purpose of the Study:

• To develop a flexible, unsupervised, and end-to-end Threat Image Projection (TIP) framework.
• To overcome the limitations of existing TIP methods by generating visually consistent and authentic synthetic X-ray images.

Main Methods:

• Introduced Meta-TIP, a novel unsupervised framework for style-adaptive Threat Image Projection.
• Developed a foreground-background contrastive loss for pure prohibited item reconstruction.
• Implemented a material-aware style-adaptive projection module for appearance control.
• Designed a logarithmic loss based on TIP principles for unsupervised optimization.

Main Results:

• Meta-TIP successfully reconstructs pure prohibited items from cluttered source images.
• The framework adaptively projects items onto target X-ray images with style consistency.
• Verified authenticity and training effectiveness of synthetic images across four public datasets (SIXray, OPIXray, PIXray, PIDray).

Conclusions:

• Meta-TIP offers a conceptually simple, flexible, and unsupervised approach to TIP.
• The framework generates highly realistic synthetic X-ray images without limitations.
• Meta-TIP significantly enhances the quality and applicability of synthetic data for security screening training.