Insulator Defect Detection in Complex Environments Based on Improved YOLOv8.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Power system safety and stability rely on effective insulator defect detection.
• Existing methods face challenges with accuracy, delay, and model size in complex environments.

Purpose of the Study:

• To develop an improved YOLOv8 target detection network for insulator defects.
• To enhance detection accuracy, reduce model size, and improve efficiency for complex environments.

Main Methods:

• Proposed an improved YOLOv8 network incorporating a C2f_DSC feature extraction module.
• Integrated EMA (encoder-modulator-attention) mechanism and BiFPN (bidirectional weighted feature pyramid network) fusion layer.
• Utilized EIOU (efficient intersection over union) loss function for accelerated convergence.

Main Results:

• Achieved a mean accuracy of 98.6% on the CPLID dataset, surpassing YOLOv8n by 0.8%.
• Reduced model size to 6.40 M, demonstrating significant efficiency.
• Outperformed other lightweight models (YOLOv8s, YOLOv6, YOLOv5s, YOLOv3Tiny) in both size and accuracy.

Conclusions:

• The proposed algorithm offers a practical and feasible solution for insulator defect detection on edge devices.
• The enhanced YOLOv8 network effectively addresses limitations of existing methods in complex scenarios.
• The study highlights the potential for improved power system safety and operational stability.