Scholarlink

<p>Precise defect detection is crucial to ensure the quality and safety of pharmaceuticals. The subtle and diverse defects on the pharmaceutical vial body pose challenges for inspection. This paper establishes a multi-station vision inspection system to capture multi-perspective imaging of the vials, develops an image processing method that combines detection area positioning and sample image segmentation, and create a custom vial body defect dataset, VBD-DET, including stain, scratch, and filling variation. We propose a Multi-scale Separable Network (MSNet) based on YOLOv8-n by introducing the module Spatially Separable Pooling Attention (SSPA) and the module Multi-scale Depthwise Feature Fusion (MDFF). The former aggregates feature information across spatial domains and improves the perception of subtle defects, the latter extracts multi-scale defects and enhances feature representation. Experiments on the VBD-DET dataset demonstrate that the proposed MSNet increases the mean Average Precision (mAP) by 11% and 15.2% compared to YOLOv11-n and YOLOv12-n, respectively. Our proposed MSNet keeps a good trade-off between detection precision and computational efficiency.</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

MSNet: an enhanced YOLOv8-based approach to detect defects in vial body

  • 期刊论文

  • 发表日期：

    2026年4月24日

查看全文
  • Haixia Xu,
  • Haohua Dai

摘要

Precise defect detection is crucial to ensure the quality and safety of pharmaceuticals. The subtle and diverse defects on the pharmaceutical vial body pose challenges for inspection. This paper establishes a multi-station vision inspection system to capture multi-perspective imaging of the vials, develops an image processing method that combines detection area positioning and sample image segmentation, and create a custom vial body defect dataset, VBD-DET, including stain, scratch, and filling variation. We propose a Multi-scale Separable Network (MSNet) based on YOLOv8-n by introducing the module Spatially Separable Pooling Attention (SSPA) and the module Multi-scale Depthwise Feature Fusion (MDFF). The former aggregates feature information across spatial domains and improves the perception of subtle defects, the latter extracts multi-scale defects and enhances feature representation. Experiments on the VBD-DET dataset demonstrate that the proposed MSNet increases the mean Average Precision (mAP) by 11% and 15.2% compared to YOLOv11-n and YOLOv12-n, respectively. Our proposed MSNet keeps a good trade-off between detection precision and computational efficiency.