A Two-Stage YOLOv5s-U-Net Framework for Defect Localization and Segmentation in Overhead Transmission Lines

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Transmission-line defect detection is vital for reliable grid operation.
• Current methods face challenges in balancing defect localization accuracy and fine segmentation detail.
• Limitations exist in handling small targets and class imbalance in defect datasets.

Purpose of the Study:

• To develop a novel cascaded two-stage framework for enhanced transmission-line defect detection and segmentation.
• To improve the precision and detail extraction of damage features in transmission lines.
• To provide an effective solution for intelligent operation and maintenance of power grids.

Main Methods:

• A cascaded framework utilizing YOLOv5s for initial defect localization and U-Net for fine segmentation.
• U-Net employs transfer learning with a VGG16 pretrained model to enhance performance with limited data.
• A hybrid loss function (Dice Loss + Focal Loss) addresses small-target detection and class imbalance.

Main Results:

• YOLOv5s achieved 87% mean Average Precision (mAP) for global localization.
• U-Net demonstrated 88% damage recognition precision.
• The framework obtained a mean Dice coefficient of 93.66% and 89% mean Intersection over Union (mIoU).

Conclusions:

• The proposed method effectively integrates target detection and fine segmentation for transmission-line defects.
• It significantly enhances detection precision and the extraction of detailed damage features.
• The framework offers a robust solution for intelligent operation and maintenance of transmission lines.