Automating electrical grid asset inspections: improving insulators health assessment with Region-Level annotations

UAV‑based remote data collection. Automating visual assessments with Computer Vision (CV) improves efficiency and consistency over manual assessments. CV models based on supervised learning have advanced performance in generic domains. However, in specific domains such as electrical grid asset inspections, annotated data is often scarce. Current annotation procedures are typically designed with a focus on asset condition reporting, which can result in a suboptimal degree of spatial detail. This work evaluates the benefits of using region‑level annotations in images of defective insulators for training state‑of‑the‑art deep learning vision models. The experiments, conducted on real‑world data, explore how increasing label detail can reduce training iterations, lower data requirements, and improve test performance. The findings highlight both the advantages and challenges of using richer annotations to support the automation of visual inspection of electrical grid assets. This enables utility companies to conduct more efficient inspections, reduce costs, and ultimately improve grid reliability.