Experimental study demonstrates improved anomaly detection in electrical insulators, indicating a cost-effective approach for automated grid inspection.
Key Points
To develop a lightweight, dual-branch visual framework that improves fine-grained insulator anomaly detection under complex outdoor backgrounds without requiring additional imaging hardware.
Constructed DBRM-Net to simultaneously process standard RGB images alongside an RGB-derived six-channel Surface Structural Modality (SSM) capturing luminance, gradient, directional, Laplacian, roughness, and local-residual descriptors.
Integrated a Prior-Aware Mixture Fusion (PAMF) module to reconcile cross-branch agreements and discrepancies, paired with an Adaptive Contextual Topology Injection Fusion (ACTIF) module to preserve RGB semantic context.
Evaluated performance against baseline detectors across two insulator anomaly datasets, including primary benchmark testing on the IFD dataset.
Achieved an AP50 of 88.7 and an AP50:95 of 65.4 on the primary IFD benchmark dataset.
Outperformed the strongest SSM-based baseline detector with modest but consistent accuracy gains while reducing parameter count and computational costs.
Demonstrated higher detector-side inference speeds and robust feature integration across repeated runs and ablation experiments.