UAV inspection insulator defect detection method based on dynamic adaptation improved YOLOv8
摘要
Insulators play an important role in ensuring the electrical stability of transmission systems, and timely detection of insulator status can effectively guarantee the safe and stable operation of power systems. However, the generalized YOLOv8 model still faces two challenges in the complex application scenario of power inspection by Unmanned Aerial Vehicles (UAVs) for high-altitude insulator defects: one is the standard convolution’s inability to efficiently extract insulator edge features, which leads to a large number of convolution operations and results in parameter redundancy; the other is the distortion of feature spatial information fusion due to complex background interference. To solve these challenges, this paper proposes a lightweight and dynamically adaptive model named DVW-YOLO (Dynamic VoV Wise YOLO, based on YOLOv8). First, a Dynamic Elliptic Convolution is designed to make the model more efficient in focusing on elliptic features, significantly reducing redundant feature extraction and multi-size fusion, and thus significantly reducing the model parameters. Then, a VoV-DE-GSCSP module is designed, and a Dynamic Adaptive Fusion Network is developed in combination with GSConv. This network restricts the cross-fusion of complex background interference information and enhances the effective feature fusion range. Finally, Wise-IoU is used to balance the performance between lightweight and average accuracy. Experimental results demonstrate that, compared with the standard YOLOv8 model, the lightweight performance of the model proposed in this paper has been greatly improved, with a reduction of 11.87% in parameter quantity and a 2.3% increase in