A lightweight multi-scale detection framework for X-ray images with supervised contrastive learning.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Security Technology

Background:

• Automated X-ray security inspection faces challenges with occluded and small objects.
• Existing object detectors struggle with dense occlusion and cluttered backgrounds in X-ray imagery.
• Limited computational resources hinder real-world applicability of current detection systems.

Purpose of the Study:

• To develop a lightweight and discriminative multi-scale object detection framework for X-ray security inspection.
• To improve the accuracy and efficiency of identifying prohibited items under challenging conditions.
• To enhance the real-world applicability of object detection in security screening.

Main Methods:

• Proposed YOLOv8-DWConv-CSAF framework integrating architectural compression, attention-guided feature enhancement, and contrastive representation learning.
• Replaced standard convolutions with depthwise separable convolutions (DWConv) for reduced parameters and computational cost.
• Introduced a novel Channel-Spatial Attention Fusion (CSAF) module and a hybrid loss function (PIoU + InfoNCE contrastive loss).

Main Results:

• YOLOv8-DWConv-CSAF achieved state-of-the-art performance on CLCXray and HiXray benchmarks.
• The model demonstrated high accuracy combined with real-time efficiency.
• Significant reduction in parameters and computational cost achieved through DWConv integration.

Conclusions:

• The proposed YOLOv8-DWConv-CSAF framework is well-suited for practical security screening systems.
• The method effectively addresses challenges of small, overlapping, and heavily occluded objects in X-ray imagery.
• Achieved robust generalization capabilities for enhanced security inspection.