Scholarlink

To ensure public safety, accurate detection of prohibited items during security inspections is crucial. These items often exhibit multi-scale variations and are difficult to accurately position, especially smaller objects in X-ray images. This article introduces the EI-YOLO algorithm, an enhanced version of YOLOv7, designed specifically for the detection of prohibited items in security contexts. The EI-YOLO algorithm incorporates several novel technological enhancements to improve detection performance. Firstly, an Adaptive Feature Module (AFM) is implemented at the end of the backbone network to aggregate multi-scale feature information effectively. Secondly, a Feature Pyramid Fusion (FPF) module integrates progressive fusion of adjacent layer features, including the B2 layer, to enhance the detection of small-scale objects and overall accuracy. Additionally, the algorithm employs the Normalized Wasserstein Distance (NWD), based on 2D Gaussian distributions, to refine bounding box regression accuracy. Evaluated on the OPIXray and HIXray datasets, EI-YOLO demonstrates superior performance compared to the baseline YOLOv7 and YOLOv8 network, confirming its efficacy in enhancing security inspection systems.

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

EI-YOLO: Efficiently Improved YOLO on Detection of Prohibited Items During Security Inspections

  • 会议论文

  • 发表日期：

    2025年

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

摘要

To ensure public safety, accurate detection of prohibited items during security inspections is crucial. These items often exhibit multi-scale variations and are difficult to accurately position, especially smaller objects in X-ray images. This article introduces the EI-YOLO algorithm, an enhanced version of YOLOv7, designed specifically for the detection of prohibited items in security contexts. The EI-YOLO algorithm incorporates several novel technological enhancements to improve detection performance. Firstly, an Adaptive Feature Module (AFM) is implemented at the end of the backbone network to aggregate multi-scale feature information effectively. Secondly, a Feature Pyramid Fusion (FPF) module integrates progressive fusion of adjacent layer features, including the B2 layer, to enhance the detection of small-scale objects and overall accuracy. Additionally, the algorithm employs the Normalized Wasserstein Distance (NWD), based on 2D Gaussian distributions, to refine bounding box regression accuracy. Evaluated on the OPIXray and HIXray datasets, EI-YOLO demonstrates superior performance compared to the baseline YOLOv7 and YOLOv8 network, confirming its efficacy in enhancing security inspection systems.