Scholarlink

<p>The reliability and safety of power transmission grids critically depend on the condition of insulators, making efficient and accurate defect detection essential. While CNN-based object detection frameworks like YOLO offer a reasonable balance between speed and accuracy, they face limitations in handling class imbalance and complex backgrounds. Transformer-based detectors (e.g., DETR) eliminate the need for Non-Maximum Suppression (NMS) but suffer from high computational costs and slower convergence. To address these challenges, the proposed method is a hybrid Detection Transformer (DETR) framework for insulator defect detection, combining semi-supervised learning with advanced loss methods. Our method introduces Stage-wise Hybrid Matching for generating high-quality pseudo-labels, focal loss to address data imbalance, and Cross-view Query Consistency to enhance feature robustness. Additionally, encoder complexity is optimized, ensuring scalability for UAV- based inspections. Experimental results demonstrate that the proposed hybrid DETR achieves 85 FPS, surpasses state-of-the-art detectors, including YOLO and DETR variants, achieving superior accuracy and speed across diverse and challenging datasets.</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

Accurate insulator defect detection in power transmission lines using semi-supervised hybrid DETR with advanced loss methods

  • 期刊论文

  • 发表日期：

    2025年9月4日

查看全文
  • Raja Sekhar Sankuri,
  • Nagesh Bhattu Sristy,
  • Sri Phani Krishna Karri

摘要

The reliability and safety of power transmission grids critically depend on the condition of insulators, making efficient and accurate defect detection essential. While CNN-based object detection frameworks like YOLO offer a reasonable balance between speed and accuracy, they face limitations in handling class imbalance and complex backgrounds. Transformer-based detectors (e.g., DETR) eliminate the need for Non-Maximum Suppression (NMS) but suffer from high computational costs and slower convergence. To address these challenges, the proposed method is a hybrid Detection Transformer (DETR) framework for insulator defect detection, combining semi-supervised learning with advanced loss methods. Our method introduces Stage-wise Hybrid Matching for generating high-quality pseudo-labels, focal loss to address data imbalance, and Cross-view Query Consistency to enhance feature robustness. Additionally, encoder complexity is optimized, ensuring scalability for UAV- based inspections. Experimental results demonstrate that the proposed hybrid DETR achieves 85 FPS, surpasses state-of-the-art detectors, including YOLO and DETR variants, achieving superior accuracy and speed across diverse and challenging datasets.