Scholarlink

<p>For the purpose of tackling the issue of sluggish identification speed and prolonged processing time resulting from the extensive network architecture of deep learning algorithms, this paper proposes Mix SwiftYOLOV5, a lightweight silicon ore recognition network based on YOLOv5, to decrease processing parameters and boost detection speed. Specifically, we introduced the GhostConv module, and using GhostConv to replace the original convolution can significantly reduce the amount of computation and the model’s size. In addition, we also introduced an innovative Mix Binary Attention (MBA) module and Spatial Dilation Pyramid (SDP) to improve network performance and feature representation capabilities. To examine the impact of the introduced method, we evaluated Mix-SwiftYOLOv5 using a self-built silicon ore dataset (a total of 4525 silicon ore images) and a PASCAL VOC2012 dataset. The experimental outcomes reveal that the proposed algorithm, when evaluated on the self-collected ore dataset, outperforms YOLOv5s by achieving a 1.61% higher mAP (0.5), a 47.5% reduction in FLOPs, a 44.2% reduction in the model parameters, and a 65.7% improvement in FPS. Through example analysis and comparison, the effectiveness and practicality of the algorithm in silicon ore detection tasks are proved.</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 lightweight and rapid method for detecting silica ores based on Yolov5

  • 期刊论文

  • 发表日期：

    2025年5月29日

查看全文
  • Jun Liu,
  • Tao Jiang,
  • Liang Guo,
  • Runjun Liu,
  • Zhihua Chen

摘要

For the purpose of tackling the issue of sluggish identification speed and prolonged processing time resulting from the extensive network architecture of deep learning algorithms, this paper proposes Mix SwiftYOLOV5, a lightweight silicon ore recognition network based on YOLOv5, to decrease processing parameters and boost detection speed. Specifically, we introduced the GhostConv module, and using GhostConv to replace the original convolution can significantly reduce the amount of computation and the model’s size. In addition, we also introduced an innovative Mix Binary Attention (MBA) module and Spatial Dilation Pyramid (SDP) to improve network performance and feature representation capabilities. To examine the impact of the introduced method, we evaluated Mix-SwiftYOLOv5 using a self-built silicon ore dataset (a total of 4525 silicon ore images) and a PASCAL VOC2012 dataset. The experimental outcomes reveal that the proposed algorithm, when evaluated on the self-collected ore dataset, outperforms YOLOv5s by achieving a 1.61% higher mAP (0.5), a 47.5% reduction in FLOPs, a 44.2% reduction in the model parameters, and a 65.7% improvement in FPS. Through example analysis and comparison, the effectiveness and practicality of the algorithm in silicon ore detection tasks are proved.