Research on insulator defect detection technology based on improved YOLOv10n

Abstract We propose an enhanced YOLOv10n algorithm to boost precision in detecting insulator defects amidst complex backgrounds during UAV-assisted inspections. The study first enhances the model’s attention to a larger range of features by labeling the existing dataset in detail, then introduces the AKConv variable convolution mechanism to the front end of the YOLOv10n backbone network, and integrates the EMA_attention multiscale attention mechanism at the end of its neural network. The enhanced YOLOv10n model achieves a 4.4% increase in mAP0.5, a 1.1% increase in mAP50-90, and a 7.1% improvement in accuracy compared to the original YOLOv10n network.