Scholarlink

<p>Remote and automatic monitoring the construction progress of pylons can effectively enhance construction efficiency and optimize costs. With the advancement of deep-learning methods in image object detection and classification, integrating these techniques with aerial monitoring in power grid engineering holds significant application value. In this study, we constructed a pylon dataset for construction progress monitoring and proposed an improved YOLO algorithm for the first time, to locate pylons in Unmanned Aerial Vehicle (UAV) aerial images and classify their construction stages. Specifically, the proposed Multi-level Resolution fusion YOLO (MRYOLO) consists of two novel modules for this task. Firstly, in order to address the shortcomings of feature pyramids in detecting large objects, we proposed a novel feature fusion network that directly integrates multi-level resolution features to optimize the gradient backpropagation path. Secondly, a lightweight attention module is embedded in the feature extraction network to enhance the global modeling capability. Furthermore, we designed a coarse-to-fine training strategy for the model. Experimental results demonstrate that MRYOLO achieved a mean average precision of 94.5% on our pylon aerial image dataset, overperforming competing models in the most accuracy metrics with real-time inference speed.</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

MRYOLO: Accurate and real-time Recognition of Pylon Construction Progress based on UAV Images

  • 期刊论文

  • 发表日期：

    2025年6月12日

查看全文
  • Xiaofeng Ma,
  • Yewei Feng,
  • Ruixi Li,
  • Liping Liu,
  • Zheng Yue,
  • Mingyue Ding

摘要

Remote and automatic monitoring the construction progress of pylons can effectively enhance construction efficiency and optimize costs. With the advancement of deep-learning methods in image object detection and classification, integrating these techniques with aerial monitoring in power grid engineering holds significant application value. In this study, we constructed a pylon dataset for construction progress monitoring and proposed an improved YOLO algorithm for the first time, to locate pylons in Unmanned Aerial Vehicle (UAV) aerial images and classify their construction stages. Specifically, the proposed Multi-level Resolution fusion YOLO (MRYOLO) consists of two novel modules for this task. Firstly, in order to address the shortcomings of feature pyramids in detecting large objects, we proposed a novel feature fusion network that directly integrates multi-level resolution features to optimize the gradient backpropagation path. Secondly, a lightweight attention module is embedded in the feature extraction network to enhance the global modeling capability. Furthermore, we designed a coarse-to-fine training strategy for the model. Experimental results demonstrate that MRYOLO achieved a mean average precision of 94.5% on our pylon aerial image dataset, overperforming competing models in the most accuracy metrics with real-time inference speed.