Research on Transmission Line Foreign Object Detection Based on Edge Calculation

Abstract

Transmission line foreign object detection plays an important role in improving the security, reliability and stability of transmission system. It's a challenge that transmission line foreign object detection achieves real-time on edge calculation as well as high performance. This paper proposes an object detector with both speed and accuracy based on YOLOv5, named YOLOv5-GHK. First, replacing Convolution, Ghost module has less parameters and calculation to achieve real-time speed. Then, network prediction with high-resolution feature improves accuracy of small transmission line foreign objects. Last, knowledge distillation ensures that lightweight model has less loss of precision. As a consequence, the proposed method runs at 36 FPS on NPU with a state-of-the-art accuracy.