Scholarlink

<p>Personal protective equipment (PPE) is pivotal in safeguarding operators in industrial and construction environments against many external physical hazards. However, prevailing PPE detection techniques frequently encounter difficulties in achieving an optimal equilibrium between detection speed and accuracy. We present the innovative Adaptive Focus YOLOv8 (AF-YOLOv8) model to address this issue, which significantly advances detection precision and speed. Firstly, we propose the dual-focused diffusion pyramid network (DFDPN), a novel architecture that masterfully captures intricate information across multiple feature scales through focusing and diffusion mechanisms. Secondly, we put forth a parameter-shared convolutional detection head (PSCDH), an efficient design that dramatically curtails the parameters required. Furthermore, we integrate the RVB_EMA module into the C2F, optimizing the model by reducing parameters while intensifying its focus on PPE characteristics. The experiments have demonstrated that AF-YOLOv8 achieves an mAP of 91.2% on the test set, representing a 1.4% improvement over the YOLOv8. Furthermore, there are 33.6%, 22%, and 34.9% reductions in parameters, GFLOPs, and model size, respectively. In addition, AF-YOLOv8 demonstrated generalization performance in experiments on the Visdrone2019 dataset.</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

AF-YOLOv8: a lightweight algorithm suitable for helmet and vest detection

  • 期刊论文

  • 发表日期：

    2025年6月9日

查看全文
  • Xi Chen,
  • Guohui Wang

摘要

Personal protective equipment (PPE) is pivotal in safeguarding operators in industrial and construction environments against many external physical hazards. However, prevailing PPE detection techniques frequently encounter difficulties in achieving an optimal equilibrium between detection speed and accuracy. We present the innovative Adaptive Focus YOLOv8 (AF-YOLOv8) model to address this issue, which significantly advances detection precision and speed. Firstly, we propose the dual-focused diffusion pyramid network (DFDPN), a novel architecture that masterfully captures intricate information across multiple feature scales through focusing and diffusion mechanisms. Secondly, we put forth a parameter-shared convolutional detection head (PSCDH), an efficient design that dramatically curtails the parameters required. Furthermore, we integrate the RVB_EMA module into the C2F, optimizing the model by reducing parameters while intensifying its focus on PPE characteristics. The experiments have demonstrated that AF-YOLOv8 achieves an mAP of 91.2% on the test set, representing a 1.4% improvement over the YOLOv8. Furthermore, there are 33.6%, 22%, and 34.9% reductions in parameters, GFLOPs, and model size, respectively. In addition, AF-YOLOv8 demonstrated generalization performance in experiments on the Visdrone2019 dataset.