Abstract: The YOLO algorithm has been widely applied to process images obtained by aerial inspection of power transmission lines. However, these UAV images often contain a large number of dense small-sized power component targets. It is difficult to detect these targets in real-time and efficiently from complex backgrounds using the YOLO algorithm alone.In this paper,we make some lightweight improvements to YOLOv8n to enhance the accuracy and speed of power component recognition. Deformable convolution sare inserted into the C2f modules of the YOLO backbone and neck,to make our method adaptively adjust the size and shape of the receptive field based on features of different scales,and obtain global and local feature representation. GSConv convolutions are integrated into the YOLO backbone,to reduce the number of model parameters and improve the detection speed. Experimental results demonstrate that our proposed method improves the recognition accuracy and speed compared to YOLOv8n, and meets the requirements of precision, lightweight, and real-time inspection of transmission line components. Specifically,the mAP50 is improved by 5.7%, the F1-score is improved by 6.0%, the number of model parameters is reduced by 7%, and the detection speed reaches 107.5 fps.

Key words: power inspection, UAV aerial photography, YOLOv8, real-time target detection, attention mechanism, deformable convolution