A Lightweight and Efficient Multi-Type Defect Detection Method for Transmission Lines Based on DCP-YOLOv8.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Electrical Engineering

Background:

• Intelligent defect detection for power line inspection is crucial for grid reliability.
• Existing AI models face a trade-off between efficiency (lightweight) and accuracy (complex) for multi-type defect identification.
• There is a need for a model that achieves high accuracy while maintaining a lightweight structure for real-time applications.

Purpose of the Study:

• To propose a lightweight and efficient multi-type defect detection method for transmission lines using AI image recognition.
• To enhance defect feature extraction and fusion for improved detection accuracy across various scales.
• To balance high detection accuracy with a reduced model size and increased processing speed.

Main Methods:

• Developed a novel method based on DCP-YOLOv8 for transmission line defect detection.
• Employed deformable convolution (C2f_DCNv3) to improve feature extraction capabilities.
• Integrated a re-parameterized cross phase feature fusion structure (RCSP) and a dynamic detection head with deformable convolution v3 (DCNv3-Dyhead) for enhanced feature expression and contextual information utilization.

Main Results:

• Achieved an average accuracy (mAP@0.5) of 72.2% on a dataset of 20 real transmission line defects, a 4.3% increase over the YOLOv8n baseline.
• Reduced model parameters to 2.8 million, a 9.15% decrease, while maintaining high detection accuracy.
• Reached a processing speed of 103 frames per second (FPS), meeting real-time detection demands.

Conclusions:

• The proposed DCP-YOLOv8 method effectively balances detection accuracy and performance for multi-type defect identification in transmission lines.
• The model demonstrates strong quantitative generalizability and meets real-time detection requirements.
• This approach offers a significant advancement in AI-powered grid transmission line inspection.