Scholarlink

<p>This paper introduces Light-YOLO, a novel lightweight framework, for efficient contraband detection in X-ray security images. It proposes three core innovations: the Channel Separation Convolution (CSC) module for enhanced feature learning, the Lightweight Shared CSC Convolutional Detection (LSCCD) head for eliminating cross-scale redundancy, and an Enhanced StarNet Backbone for hardware-aligned optimization. Light-YOLO is designed with multiple scalable variants (n, s, m, l), offering flexible deployment options across diverse hardware platforms. Extensive evaluations on the challenging OPIXray dataset demonstrate Light-YOLO’s superior efficiency–accuracy trade-off. Our models achieve substantial reductions in parameter count and GFLOPs across all scales compared to YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12. Crucially, Light-YOLO maintains highly competitive detection accuracy on OPIXray. For instance, Light-YOLON achieves the lowest GFLOPs (3.8) and parameters (1.6M) among all tested models with competitive accuracy, while Light-YOLOS delivers higher mAP50-95 (0.413) at significantly lower computational cost (15 GFLOPs) than YOLOv10s and YOLOv11s. These results affirm Light-YOLO’s practical viability for real-world, resource-constrained security inspection systems, highlighting the importance of synergistic backbone–neck–head co-design.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Light-YOLO: a lightweight framework for multi-scale contraband detection in X-ray security images via channel-decoupled feature learning

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Ran Wang,
  • Yang Zhou,
  • Guanghuan Hu,
  • Xianghua Xu

摘要

This paper introduces Light-YOLO, a novel lightweight framework, for efficient contraband detection in X-ray security images. It proposes three core innovations: the Channel Separation Convolution (CSC) module for enhanced feature learning, the Lightweight Shared CSC Convolutional Detection (LSCCD) head for eliminating cross-scale redundancy, and an Enhanced StarNet Backbone for hardware-aligned optimization. Light-YOLO is designed with multiple scalable variants (n, s, m, l), offering flexible deployment options across diverse hardware platforms. Extensive evaluations on the challenging OPIXray dataset demonstrate Light-YOLO’s superior efficiency–accuracy trade-off. Our models achieve substantial reductions in parameter count and GFLOPs across all scales compared to YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12. Crucially, Light-YOLO maintains highly competitive detection accuracy on OPIXray. For instance, Light-YOLON achieves the lowest GFLOPs (3.8) and parameters (1.6M) among all tested models with competitive accuracy, while Light-YOLOS delivers higher mAP50-95 (0.413) at significantly lower computational cost (15 GFLOPs) than YOLOv10s and YOLOv11s. These results affirm Light-YOLO’s practical viability for real-world, resource-constrained security inspection systems, highlighting the importance of synergistic backbone–neck–head co-design.