FSVM: A Few-Shot Threat Detection Method for X-ray Security Images.

Area of Science:

• Computer Vision
• Machine Learning
• Security Systems

Background:

• Automatic threat detection in X-ray baggage is crucial for security.
• Training detectors requires extensive annotated data, especially for rare contraband.
• Existing methods struggle with limited labeled samples for novel threats.

Purpose of the Study:

• To propose a few-shot threat detection model (FSVM) for identifying unseen contraband with minimal data.
• To improve the accuracy and efficiency of threat detection in X-ray security.
• To address the challenge of data scarcity for rare contraband items.

Main Methods:

• Developed a few-shot SVM-constraint (FSVM) model.
• Integrated a derivable SVM layer for back-propagation of supervised information.
• Utilized a combined loss function with SVM loss as an additional constraint.
• Evaluated on the SIXray dataset using 10-shot and 30-shot scenarios.

Main Results:

• FSVM demonstrated superior performance compared to four common few-shot detection models.
• Achieved high accuracy in detecting unseen contraband with limited samples.
• Showcased suitability for complex, distributed datasets like X-ray baggage.

Conclusions:

• FSVM is an effective approach for few-shot threat detection in X-ray baggage.
• The model overcomes data limitations for rare contraband detection.
• FSVM offers a promising solution for enhancing security inspection systems.