Answers are grounded in this paper and cite their sources.
The intelligent inspection of power grids, a critical component of modern internet of things (IoT) ecosystems, increasingly relies on embodied intelligent agents like drones to ensure operational safety. However, the perceptual accuracy of these agents in identifying insulator anomalies is often hindered by the high cost of acquiring fine-grained, pixel-level annotations for training. To address this issue, this paper proposes an advanced, weakly supervised computation method for the precise localization of insulator anomaly regions, designed for deployment on such embodied agents. Our approach requires only bounding box annotations. Specifically, an object detection network first provides coarse localization of anomalies. Subsequently, a novel multi-view cropping strategy generates diverse visual inputs to enhance local feature representation. These views are then processed by the Segment Anything Model, and a multi-view fusion strategy is employed to considerably improve the accuracy of boundary details. Experiments on a self-built insulator anomaly dataset show that our method achieves localization performance close to fully supervised approaches while substantially reducing annotation costs. This work provides a robust and efficient intelligent computation solution that enhances the perceptual capabilities of embodied agents, demonstrating considerable engineering value for autonomous inspection within smart grid IoT systems.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: