Applicability Analysis of LSK and P2 Fusion in YOLOv11 for Insulator Defect Instance Segmentation

Insulator defect instance segmentation in unmanned aerial vehicle (UAV)-based power inspection scenarios remains challenging because of large target-scale variation, complex backgrounds, weak defect textures, and limited annotated samples. To examine whether common structural enhancement strategies can improve performance in this small-sample setting, this study investigates the applicability of two modifications to YOLOv11-seg: introducing a Large Selective Kernel (LSK) module into deep backbone stages and incorporating a P2 high-resolution feature map into the feature fusion network. Experiments were conducted on an expanded insulator defect instance segmentation dataset containing 836 images, including 138 images with defect instances. To reduce the influence of a single random partition, three independent stratified data splits were constructed, and all results were reported as mean ± standard deviation across the three splits. The results show that, within the YOLOv11-seg framework, none of the LSK-based, P2-based, or LSK+P2 variants provides a clear and consistent improvement over the baseline. Although some variants achieve slightly higher mean values in individual box-level metrics, the differences remain within the range of split-to-split variation and do not support a robust performance advantage. In addition, external comparisons with Mask R-CNN, pretrained YOLOv8s-seg, and pretrained YOLOv11s-seg provide a broader reference for the performance level of different instance segmentation frameworks under the current setting. The results show that YOLOv11s-seg remains competitive among YOLO-family models, while YOLOv8s-seg achieves slightly higher average performance. These findings suggest that increasing structural complexity does not necessarily lead to robust performance gains in small-sample and class-imbalanced insulator defect instance segmentation and that the practical value of structural modifications should be evaluated cautiously under repeated data splits.