A lightweight YOLOv7 insulator defect detection algorithm based on DSC-SE.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Electrical Engineering

Background:

• Transmission line insulator inspection is crucial for power grid reliability.
• Existing Unmanned Aerial Vehicle (UAV) based detection methods often suffer from high complexity and slow speeds.
• Accurate and efficient insulator defect detection is essential for preventing failures.

Purpose of the Study:

• To develop a lightweight and efficient insulator defect detection algorithm for UAV applications.
• To address the limitations of high model complexity and inferior detection speed in current methods.
• To enhance the real-time performance of insulator defect detection systems.

Main Methods:

• A lightweight YOLOv7 algorithm incorporating Depthwise Separable Convolution (DSC) with SE channel attention (DSC-SE) in the backbone.
• Utilizing Grid Sensitive Convolution (GSConv) for feature fusion to reduce parameters and computational load.
• Implementing Efficient-IoU (EIoU-loss) in the prediction head for faster model convergence.

Main Results:

• The improved model achieved a 95.2% recognition accuracy with a model size of 7.9M.
• Achieved a 54.5% reduction in GFLOPs and a 37.8% compression in model size compared to YOLOv7.
• Demonstrated a single image detection time of 105ms and a capture rate of 13FPS on Jetson Nano.

Conclusions:

• The proposed lightweight YOLOv7 algorithm effectively balances accuracy and detection speed for real-time insulator defect inspection.
• The optimized model meets the practical demands of UAV-based power grid maintenance.
• The integration of DSC-SE, GSConv, and EIoU-loss offers a significant advancement in efficient computer vision for infrastructure monitoring.