Scholarlink

<p>Pharmaceutical blister packs face challenges such as missing capsules, incorrect capsules, and defects during encapsulation. The current deep learning-based visual inspection system in high-speed inspection scenarios faces the problem of difficulty in balancing detection accuracy and inference speed. To address this issue, this study proposes ADF-YOLO, an improved object detection model designed to address the performance limitations of conventional YOLO architectures in high-speed drug detection scenarios. The network’s fundamental computational structure is enhanced by replacing standard convolution layers with dynamic convolution operators, thereby significantly optimizing feature representation. Furthermore, the original backbone network’s A2C2f module is replaced with the FasterNet module, limiting model complexity while enhancing the extraction of critical channel information. Experimental results demonstrate that ADF-YOLO significantly outperforms the YOLOv12m baseline in industrial product defect detection, achieving a 9.8% improvement in mAP (from 79.5% to 92.3%). The model also exhibits a 33.5% faster inference speed while decreasing the parameter count by 49.6%. Additional validation on the RoboFlow public dataset confirms the algorithm's strong generalization capability across different domains. Therefore, ADF-YOLO provides a practical and efficient technical pathway for real-time online inspection, particularly suited to pharmaceutical packaging quality control applications.</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

ADF-YOLO: lightweight adaptive network for high-speed pharmaceutical blister pack detection

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Yifu Yu,
  • Peng Zhou,
  • Haikang Fu,
  • Chunbing Tang

摘要

Pharmaceutical blister packs face challenges such as missing capsules, incorrect capsules, and defects during encapsulation. The current deep learning-based visual inspection system in high-speed inspection scenarios faces the problem of difficulty in balancing detection accuracy and inference speed. To address this issue, this study proposes ADF-YOLO, an improved object detection model designed to address the performance limitations of conventional YOLO architectures in high-speed drug detection scenarios. The network’s fundamental computational structure is enhanced by replacing standard convolution layers with dynamic convolution operators, thereby significantly optimizing feature representation. Furthermore, the original backbone network’s A2C2f module is replaced with the FasterNet module, limiting model complexity while enhancing the extraction of critical channel information. Experimental results demonstrate that ADF-YOLO significantly outperforms the YOLOv12m baseline in industrial product defect detection, achieving a 9.8% improvement in mAP (from 79.5% to 92.3%). The model also exhibits a 33.5% faster inference speed while decreasing the parameter count by 49.6%. Additional validation on the RoboFlow public dataset confirms the algorithm's strong generalization capability across different domains. Therefore, ADF-YOLO provides a practical and efficient technical pathway for real-time online inspection, particularly suited to pharmaceutical packaging quality control applications.