An Improved YOLOv8n Network for Transmission Line Intrusion Foreign Object Detection

To address the problems of low foreign object detection accuracy and high model complexity due to many small targets and complex surrounding environment in transmission line scenarios, a foreign object detection network CFG-YOLOv8n is proposed based on YOLOv8n. Firstly, a CA mechanism is introduced to increase the attention of YOLOv8n to occluded targets and small targets, so as to increase the detection accuracy. Secondly, ghost convolution is employed to replace conventional convolution in the backbone and neck networks to reduce the model parameter amount. Finally, Focal-EIoU loss is introduced to alleviate the sample imbalance problem during bounding box regression, and increase foreign object detection accuracy. Experiments demonstrate that the mAP of the proposed CFG-YOLOv8n network is 97.6%, and the size and the parameter quantity of the model are 5.51MB and 3.1M, respectively. Compared with YOLOv8n model, the mAP is increased by 1%, and the model size and the parameter quantity are reduced by 7% and 3%, respectively. Moreover, the CFG-YOLOv8n model exhibits lower missed detection rates for occluded and small targets, which has certain practical significance for ensuring the safety of transmission lines.