MS-YOLO: An Improved Lightweight Transmission Line Insulator Defect Detection Algorithm Based on YOLOv8

The overhead transmission line insulators are more prone to defects as a result of prolonged exposure to the ever-changing outdoor conditions. Currently, the mainstream method is to use drones for aerial photography, but the weak computing power and low image quality of drones result in lower accuracy in defect detection. In this regard, this paper proposes an improved lightweight YOLOv8 insulator defect detection algorithm (MS-YOLO). Firstly, the MHSA (Multi-headed Self-attention) is implemented to augment the focus of model on insulators and their defects in complex backgrounds. Secondly, the YOLOv8 backbone network is modified by replacing the C2f convolutional layer with C2f_DCN to reduce the focus of the feature extraction layer on non-insulator image regions, further strengthening the recognition ability. Finally, to expedite the network’s convergence rate and reduce its loss magnitude, the model opts for SIoU (SCYLLA-IoU) Loss as its loss function. It is evidenced by experimental results that in comparison with the baseline model, accuracy is boosted by 1.5% and FPS is boosted by 20%. While the accuracy is slightly improved, the speed of small target detection is greatly improved, providing a reference for insulator defect detection.