Scholarlink

<p>Existing transmission line insulator diagnostics confront a tripartite challenge: undetected sub-pixel anomalies, interference from heterogeneous backgrounds, and insufficient computational headroom at edge nodes. To address these issues, this study proposes a lightweight detection model named TRS-YOLO, based on YOLOv12. In the TRS-YOLO model, we introduce three key innovations: (1) a receptive field attention convolution module (RFAConv) is proposed to enhance the feature extraction capability of the backbone network; (2) a spatial and channel synergistic attention module (SCSA) is integrated, effectively boosting the model's perception of critical insulator features—particularly improving robustness for small defect detection in complex backgrounds—through a synergistic mechanism combining multi-semantic spatial attention guidance and progressive channel self-attention recalibration; (3) to enhance the detection head's sensitivity to subtle defects while maintaining model efficiency, a mobile inverted bottleneck block (MBConv) is innovatively introduced. Validated on a dataset where small targets constitute&#xa0;63.96% of samples, TRS-YOLO achieves&#xa0;90.7% mAP@50—surpassing YOLOv12 by 2.9%&#xa0;and&#xa0;outperforming state-of-the-art YOLOv13 by 4.2%—all while maintaining&#xa0;ultra-efficient deployment&#xa0;with only&#xa0;2.95&#xa0;M parameters&#xa0;and&#xa0;166 FPS real-time inference speed. This demonstrates TRS-YOLO’s unique capability to deliver&#xa0;superior accuracy for critical small targets&#xa0;and&#xa0;exceptional lightweight adaptability&#xa0;for edge-device deployment.</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

TRS-YOLO: a lightweight insulator defect detection method based on enhanced YOLOv12

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Jiao Wang,
  • Bin Li

摘要

Existing transmission line insulator diagnostics confront a tripartite challenge: undetected sub-pixel anomalies, interference from heterogeneous backgrounds, and insufficient computational headroom at edge nodes. To address these issues, this study proposes a lightweight detection model named TRS-YOLO, based on YOLOv12. In the TRS-YOLO model, we introduce three key innovations: (1) a receptive field attention convolution module (RFAConv) is proposed to enhance the feature extraction capability of the backbone network; (2) a spatial and channel synergistic attention module (SCSA) is integrated, effectively boosting the model's perception of critical insulator features—particularly improving robustness for small defect detection in complex backgrounds—through a synergistic mechanism combining multi-semantic spatial attention guidance and progressive channel self-attention recalibration; (3) to enhance the detection head's sensitivity to subtle defects while maintaining model efficiency, a mobile inverted bottleneck block (MBConv) is innovatively introduced. Validated on a dataset where small targets constitute 63.96% of samples, TRS-YOLO achieves 90.7% mAP@50—surpassing YOLOv12 by 2.9% and outperforming state-of-the-art YOLOv13 by 4.2%—all while maintaining ultra-efficient deployment with only 2.95 M parameters and 166 FPS real-time inference speed. This demonstrates TRS-YOLO’s unique capability to deliver superior accuracy for critical small targets and exceptional lightweight adaptability for edge-device deployment.