Scholarlink

<p>Accurate and efficient algorithms for detection are essential technologies utilized by robots for plum harvesting. To facilitate effective implementation on edge computing devices that have restricted processing capabilities, this research introduces a lightweight plum detection algorithm derived from YOLOv5s. Initially, the backbone network of the original architecture is substituted with the Revcol network. Subsequently, the feature space pyramid pooling (SPPF) is refined by integrating a Large Separable Kernel Attention (LSKA) mechanism, and the C3 module of the original model is improved through the use of Distributed Shift Convolutions (DSConv). Finally, a custom-designed lightweight detection head is employed to construct the lightweight plum detection model. The results show that the average mAP of the improved detection model is 96.3%, and the number of floating point operations (FLOPs) and parameters are reduced to 84.8% and 91.2% respectively. On edge computing devices, the inference speed reaches 44.64 frames per millisecond. This algorithm features high accuracy, lightweight design, and fast inference speed, providing valuable references for real-time object detection deployment of plum harvesting robots on edge computing devices.</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 real-time algorithm for plum harvesting detection in orchards under complex conditions

  • 期刊论文

  • 发表日期：

    2025年2月22日

查看全文
  • Dongyan Zhang,
  • Nuo Chen,
  • Siyu Mao,
  • Chenxv Wu,
  • Dandan Gu,
  • Lanxiang Zhang

摘要

Accurate and efficient algorithms for detection are essential technologies utilized by robots for plum harvesting. To facilitate effective implementation on edge computing devices that have restricted processing capabilities, this research introduces a lightweight plum detection algorithm derived from YOLOv5s. Initially, the backbone network of the original architecture is substituted with the Revcol network. Subsequently, the feature space pyramid pooling (SPPF) is refined by integrating a Large Separable Kernel Attention (LSKA) mechanism, and the C3 module of the original model is improved through the use of Distributed Shift Convolutions (DSConv). Finally, a custom-designed lightweight detection head is employed to construct the lightweight plum detection model. The results show that the average mAP of the improved detection model is 96.3%, and the number of floating point operations (FLOPs) and parameters are reduced to 84.8% and 91.2% respectively. On edge computing devices, the inference speed reaches 44.64 frames per millisecond. This algorithm features high accuracy, lightweight design, and fast inference speed, providing valuable references for real-time object detection deployment of plum harvesting robots on edge computing devices.