A Novel Dangerous Goods Detection Network Based on Multi-Layer Attention Mechanism in X-Ray Baggage Images

Aim

        To develop an improved dangerous goods detection system using YOLOv11 that achieves higher accuracy and real-time performance in identifying prohibited items in X-ray baggage images.

Abstract

        This project proposes an enhanced dangerous goods detection system for X-ray baggage security inspection using YOLOv11. While existing methods like ResNet-50 and YOLOv5 have shown promising results, they suffer from accuracy limitations in detecting overlapping and occluded objects. Our proposed system leverages YOLOv11’s advanced architecture to detect 12 classes of dangerous items (Baton, Bullet, Gun, Hammer, Handcuffs, Knife, Lighter, Pliers, Powerbank, Scissors, Sprayer, Wrench) with improved precision and faster inference speed, making it suitable for real-time security screening applications.

Proposed System

       Implementation of YOLOv11 for dangerous goods detection with:

• 12 Classes: Baton, Bullet, Gun, Hammer, Handcuffs, Knife, Lighter, Pliers, Powerbank, Scissors, Sprayer, Wrench
• Advanced feature pyramid network
• Optimized anchor-free detection head
• Enhanced data augmentation techniques
• Transfer learning from pre-trained weights
• Real-time inference capability (>30 FPS)

         The proposed system implements YOLOv11, the latest version of the YOLO object detection architecture, optimized for detecting dangerous goods in X-ray baggage images. Unlike the base paper’s ResNet-50 and YOLOv5 approaches, YOLOv11 offers improved feature extraction through C3k2 modules and enhanced spatial attention mechanisms. The system detects 12 threat classes: Baton, Bullet, Gun, Hammer, Handcuffs, Knife, Lighter, Pliers, Powerbank, Scissors, Sprayer, and Wrench. Key improvements include faster training convergence, better real-time inference speed, and better handling of occluded objects through advanced feature pyramid networks. The anchor-free detection head and dynamic task-aligned assigner enable precise localization of dangerous items even in cluttered baggage scenarios.

Advantages

1. Higher Accuracy: Improved mAP through YOLOv11’s advanced architecture
2. Faster Training: Reduced epochs required due to better convergence
3. Real-time Detection: Optimized for speed without sacrificing accuracy
4. Better Generalization: Handles occlusion and overlapping objects effectively
5. Lower False Positives: Enhanced feature extraction reduces noise
6. Scalability: Easily deployable across multiple security checkpoints