Insulators' Identification and Missing Defect Detection in Aerial Images Based on Cascaded YOLO Models

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Intelligent inspection of high-voltage transmission lines requires accurate insulator identification and defect detection.
• Complex backgrounds, occlusion, and small defect sizes pose challenges for aerial image analysis.

Purpose of the Study:

• To develop and evaluate cascaded You Only Look Once (YOLO) models for insulator identification and missing defect detection.
• To improve upon existing methods for detecting small, occluded defects in challenging aerial imagery.

Main Methods:

• Creation of specialized datasets for insulator location and missing defect detection.
• Development of an improved YOLOv3-dense based model for insulator localization.
• Utilization of an enhanced YOLOv4-tiny model for missing defect detection on identified insulators.
• Cascading the YOLO models for a comprehensive identification and defect detection system.

Main Results:

• The cascaded YOLO models achieved an average precision of 98.4% for missing defect detection.
• This represents a significant improvement over Faster R-CNN (5.2%) and SSD (10.2%).
• The system operates at a high speed of 106 frames per second.

Conclusions:

• The proposed deep learning models demonstrate high performance for insulator identification and missing defect detection.
• This technology enhances the safety and efficiency of high-voltage transmission line inspections.
• The cascaded YOLO approach effectively addresses challenges posed by complex backgrounds and small defects.