Scholarlink

<p>Bird intrusion around transmission lines increases the risk of insulator contamination and line faults. Rapid and reliable on-site monitoring is therefore becoming increasingly important. In practical deployments, edge devices often impose strict computational budgets. Detection must remain accurate while being lightweight and real time. This paper proposes TLIBDNet, a lightweight bird-intrusion detector designed for transmission-line scenarios. The method is built on a YOLOv11-style one-stage detector. It reduces redundant computation by replacing the backbone with MobileNetV3. To improve robustness to scale variation and small distant birds, the network introduces an AFPN-based multi-scale feature fusion module. This module strengthens interactions between semantic and fine-grained features. We also build a transmission-line intrusion bird dataset for training bird recognition models. The dataset contains 9,585 images. Experiments show that TLIBDNet achieves a mean average precision of 81.4 at an IoU of 0.5. It requires 2.21&#xa0;billion floating-point operations and processes 28 frames per second on a Jetson edge platform. Compared with two-stage detectors, transformer-based detectors, and multiple YOLO-series baselines, TLIBDNet provides a better balance of accuracy, speed, and resource cost for on-site deployment in bird-hazard monitoring and repelling systems.</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

Improved YOLOv11-based algorithm for identifying birds intruding on power lines

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月19日

查看全文
  • Jingliang Song,
  • Yubin She,
  • Yongliang Li,
  • Jia Yang,
  • Lin Liu,
  • Chen Fu,
  • Hongyu Di

摘要

Bird intrusion around transmission lines increases the risk of insulator contamination and line faults. Rapid and reliable on-site monitoring is therefore becoming increasingly important. In practical deployments, edge devices often impose strict computational budgets. Detection must remain accurate while being lightweight and real time. This paper proposes TLIBDNet, a lightweight bird-intrusion detector designed for transmission-line scenarios. The method is built on a YOLOv11-style one-stage detector. It reduces redundant computation by replacing the backbone with MobileNetV3. To improve robustness to scale variation and small distant birds, the network introduces an AFPN-based multi-scale feature fusion module. This module strengthens interactions between semantic and fine-grained features. We also build a transmission-line intrusion bird dataset for training bird recognition models. The dataset contains 9,585 images. Experiments show that TLIBDNet achieves a mean average precision of 81.4 at an IoU of 0.5. It requires 2.21 billion floating-point operations and processes 28 frames per second on a Jetson edge platform. Compared with two-stage detectors, transformer-based detectors, and multiple YOLO-series baselines, TLIBDNet provides a better balance of accuracy, speed, and resource cost for on-site deployment in bird-hazard monitoring and repelling systems.