STID: A Two-Stage Cascade Framework for Defect-Annotation-Free Insulator Defect Detection

doi:10.1109/tim.2026.3709419

STID: A Two-Stage Cascade Framework for Defect-Annotation-Free Insulator Defect Detection

2026-07-02

0

PRE

AI

DOI：10.1109/tim.2026.3709419

Original

Original request for help

Save

Abstract

En 中文

Insulators are critical components in overhead transmission lines, where long-term environmental exposure can cause subtle defects that compromise grid reliability. In practice, defect annotations are scarce, and images are often dominated by complex backgrounds, making defect cues weak and unstable when learned directly from raw scenes. This article proposes STID, a practical two-stage framework for insulator inspection that decouples background suppression from anomaly learning. Stage I is weakly supervised with coarse insulator masks to extract insulator-only regions, constructing an insulator-centric domain that suppresses background clutter and reduces background-induced variance. The extracted instances are normalized into square patches and processed in Stage II by a revised normal-only anomaly detector based on reverse distillation, termed SDS-Net, which is enhanced with an SCFB module for structure-aware feature bridging and a dynamic intensity mechanism for adaptive anomaly response modulation, together with an inference-time score refinement scheme for more robust anomaly scoring. The detector produces pixel-level anomaly maps and instance-level anomaly scores without requiring defect labels or defect masks for anomaly learning. Experiments on the DINS dataset demonstrate that STID achieves 93.13 % image-level AUROC (I-AUROC) and 96.57 % pixel-level AUROC (P-AUROC), indicating reliable defect detection and localization under cluttered backgrounds. The proposed framework provides a practical solution for insulator defect inspection and may be extended to other inspection scenarios involving repetitive structures and strong background interference. The implementation is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/EEML-HC/STID</uri>

Keywords:

Anomaly detection (AD)

background suppression

Insulator defect detection

normal-only learning

weak supervision

STID: A Two-Stage Cascade Framework for Defect-Annotation-Free Insulator Defect Detection

Abstract

Journal

Organization