Hierarchical Distillation of SAM for Lightweight Insulator Defect Segmentation

The real-time detection of insulator failures is particularly critical for the safety of power transmission networks. Traditional drone inspections are often constrained by limited computational resources and network bandwidth, making the real-time detection of defects an increasingly pressing requirement. This paper introduces a novel approach for insulator defect segmentation by hierarchically distilling a Segment Anything Model (SAM) to develop a lightweight UNet model. We propose utilizing defect detection boxes as segmentation prompts, thereby equipping SAM with the capability for defect segmentation. Through a phased distillation process, we refine a lightweight UNet model that maintains high accuracy while significantly reducing computational load. Specifically, we introduce a hierarchical distillation strategy that encompasses feature embedding distillation and segmentation mask distillation. Extensive experiments validate the efficiency and applicability of the proposed model for high-voltage insulator defect segmentation.