Star-IBR: Star-Shaped Iterative Box Refinement for Insulator Defect Localization

Accurate localization of insulator defects remains challenging in transmission-line inspection, particularly for single-stage detectors evaluated at stringent overlap thresholds. Evaluation can also be biased when transformed views of one scene are divided between training and test subsets. We propose Star-Shaped Iterative Box Refinement (Star-IBR) to improve localization accuracy and introduce a scene-disjoint evaluation protocol for IDD-S22. Star-IBR decodes a distribution-regressed box, samples nine boundary-aware locations, and predicts an additive correction; stop-gradient blocks direct refined-loss paths to initial coordinates. Reconstructing scene identities showed that all 141 scenes in the official test subset occur in training. Under scene-disjoint evaluation, Star-IBR increased average precision (AP) over intersection-over-union thresholds 0.50–0.95 from 0.529 to 0.556 relative to the distribution-regression baseline across independent five-seed sweeps. The mean gain was 0.0268 (95% confidence interval, 0.0192–0.0344), with a gain of 0.0264 at a threshold of 0.75. Paired five-seed contrasts across sampling geometry, initialization, and gradient routing ranged from −0.0016 to +0.0036 AP; all 95% confidence intervals included zero. The head increased parameter count by 0.41% and V100 latency by 4.7% while reducing throughput by 4.5%. On IDD-S22, Star-IBR improves localization at stringent IoU thresholds, while scene-disjoint partitioning provides a reproducible evaluation protocol without cross-subset scene overlap.