Unattended Substation Inspection Algorithm Based on Improved YOLOv5

Abstract

The lack of detection accuracy has been the pain point of unattended substation inspection at all times. One detection algorithm in terms of the improved YOLOv5 is proposed in the paper so as to enhance the detection accuracy. A backbone with unique attention mechanism is designed to extract more accurate feature maps. The improved backbone increases the sensitivity of the model to channel features by accurately location information relations and long-range dependencies with a long range are encoded together with a spatial direction as well as accurate location information with the other one is preserved, helping the algorithm to locate inspection objects. The coming results through experiments demonstrate the detection algorithm containing the SE attention has 0.7% improvement on mAP, while the detection algorithm containing the CA has 1.3% improvement on mAP, and the detection algorithm containing CA is more suitable for unattended substation inspection.