Scholarlink

<p>To address the slow speed, low accuracy, high computational demands, and deployment challenges on industrial edge devices in coal gangue detection, we propose a lightweight improved YOLOv11 algorithm. Key innovations include: (1) replacing the backbone with ShuffleNetV2 to accelerate inference, (2) introducing lightweight ADown sampling to reduce model complexity while increasing average detection accuracy, (3) enhancing the C2PSA module by integrating Triplet Attention, creating the C2PSA-TriAtt module to strengthen multi-dimensional feature focus, and (4) proposing the Inner-FocalerIoU loss function to replace CIoU for optimized bounding box regression. Experiments demonstrate 99.10% detection accuracy. The model reduces size by 38%, parameters by 41%, computation by 40%, and per-image detection time by 1&#xa0;ms. These improvements significantly boost speed and accuracy, enabling practical deployment on edge devices. Critically, the model also reduces false detections caused by surface coal dust, supporting more efficient coal processing and resource utilization in industrial settings.</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

SNAT-YOLO: Efficient Cross-Layer Aggregation Network for Edge-Oriented Gangue Detection

  • 期刊论文

  • 发表日期：

    2025年11月17日

查看全文
  • Shang Li,
  • Yuan Liu,
  • Zeyu Tang

摘要

To address the slow speed, low accuracy, high computational demands, and deployment challenges on industrial edge devices in coal gangue detection, we propose a lightweight improved YOLOv11 algorithm. Key innovations include: (1) replacing the backbone with ShuffleNetV2 to accelerate inference, (2) introducing lightweight ADown sampling to reduce model complexity while increasing average detection accuracy, (3) enhancing the C2PSA module by integrating Triplet Attention, creating the C2PSA-TriAtt module to strengthen multi-dimensional feature focus, and (4) proposing the Inner-FocalerIoU loss function to replace CIoU for optimized bounding box regression. Experiments demonstrate 99.10% detection accuracy. The model reduces size by 38%, parameters by 41%, computation by 40%, and per-image detection time by 1 ms. These improvements significantly boost speed and accuracy, enabling practical deployment on edge devices. Critically, the model also reduces false detections caused by surface coal dust, supporting more efficient coal processing and resource utilization in industrial settings.