Balanced X-ray Security Dataset and Enhanced YOLO for Contraband Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Technology

Background:

• Existing X-ray contraband detection datasets suffer from severe class imbalance and limited high-quality annotated data.
• Poor model adaptability to complex scenarios hinders effective contraband detection in real-world applications.

Purpose of the Study:

• To construct a balanced X-ray contraband detection dataset addressing class imbalance and data scarcity.
• To develop lightweight and high-precision detection models adaptable to complex scenarios.

Main Methods:

• A balanced dataset of 13,728 images across 12 contraband categories was created by combining SIXray and PIDray datasets.
• A Class-Specific Augmentation Framework (CSAF) and random undersampling were employed to achieve uniform class distribution.
• Two improved YOLOv11s-based models, ASEA-Net and CSEC-Net, were proposed for enhanced detection.

Main Results:

• ASEA-Net achieved 95.78% accuracy and 93.55% mAP@50, outperforming YOLOv11s with fewer parameters.
• CSEC-Net significantly reduced parameters and FLOPs, enabling deployment on resource-constrained edge devices.
• Both models demonstrated strong performance in complex scenarios.

Conclusions:

• The balanced dataset effectively addresses class imbalance and enhances model training for X-ray contraband detection.
• ASEA-Net and CSEC-Net offer improved accuracy, efficiency, and adaptability for security screening applications.
• The proposed models validate the value of balanced datasets and novel architectures in advanced contraband detection systems.