Scholarlink

<p>Surface defect detection in filled vials is significant for the pharmaceutical safety. Due to the weak features and different scales of defects, it leads to missed and false inspection. In this paper, we firstly design a data acquisition solution and create the custom datasets VialG1_DET, VialG2_DET, VialG3_DET. Secondly, we design a multi-workstation inspection system combining traditional image processing algorithms and deep learning object detection algorithms to detect defects of surface and contents in vials. We propose Defect Detection of Surface and Contents in Vials (<b>DDSCNet</b>) by designing Quadra Fusion and Attention (QUFUAtt) module which enhances the capability of feature fusion in network, introducing the self-attention and convolution (ACmix) which focuses on the defective areas, and Linear Deformable Convolution which extracts the weak features of defects. Our experiments show that the proposed <b>DDSCNet</b> achieves 76.7% mean Average Precision (mAP@0.5) on the VialG1_DET, 65.9% mAP@0.5 on the VialG2_DET, along with 86.9% mAP@0.5 on the VialG3_DET with low computational complexity of 9.3GFLOPS, and outperforms YOLOv11 by 3.5% mAP@0.5.</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

A vision-based inspection system for pharmaceutical production line

  • 期刊论文

  • 发表日期：

    2025年3月14日

查看全文
  • Haixia Xu,
  • Yuting Xu,
  • Kaiyu Hu

摘要

Surface defect detection in filled vials is significant for the pharmaceutical safety. Due to the weak features and different scales of defects, it leads to missed and false inspection. In this paper, we firstly design a data acquisition solution and create the custom datasets VialG1_DET, VialG2_DET, VialG3_DET. Secondly, we design a multi-workstation inspection system combining traditional image processing algorithms and deep learning object detection algorithms to detect defects of surface and contents in vials. We propose Defect Detection of Surface and Contents in Vials (DDSCNet) by designing Quadra Fusion and Attention (QUFUAtt) module which enhances the capability of feature fusion in network, introducing the self-attention and convolution (ACmix) which focuses on the defective areas, and Linear Deformable Convolution which extracts the weak features of defects. Our experiments show that the proposed DDSCNet achieves 76.7% mean Average Precision (mAP@0.5) on the VialG1_DET, 65.9% mAP@0.5 on the VialG2_DET, along with 86.9% mAP@0.5 on the VialG3_DET with low computational complexity of 9.3GFLOPS, and outperforms YOLOv11 by 3.5% mAP@0.5.