HDSI-YOLO: Hierarchical-semantics Driven Insulator Defect Detection with Cross-domain Generalization

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Insulator defect detection is a core task of intelligent inspection of power transmission lines, but existing deep-learning detectors face two bottlenecks: severe semantic ambiguity among rare defect classes (broken vs. pollution-flashover) under flat classification heads, and poor cross-domain generalization across datasets with different imaging conditions and class taxonomies. This paper proposes HDSI-YOLO, a hierarchical-semantics-driven detector that injects the well-known-but-never-modeled class hierarchy of insulator defects into a YOLOv8 base: (i) a hierarchical detection head (HierarchicalDetect) that adds a lightweight parent-class branch parallel to the decoupled head; (ii) a hierarchical supervision loss (ChildMapLoss) that derives parent labels automatically from child labels via a mapping table at zero annotation cost; and (iii) a defect feature amplification module (DFAM) with a P2 high-resolution branch. On CPLID single-domain ablations, HDSI-YOLO improves mAP50-95 by +0.009 and defect AP by +0.020, with the hierarchical head and loss as the main gain source. Under joint CPLID+IDID training with 3-seed paired tests, gains on the original official IDID validation split appeared significant (mAP50-95 +1.42 pp, p=0.014); however, a subsequent split audit showed that this split contains scene-level leakage (41.7% of images share source scenes with the training set), and on a leakage-free evaluation set the gains lose significance (+0.75 pp, p=0.301), with only a small nominal flashover gain remaining (+0.17 pp, p=0.022). We report both evaluation settings transparently. The paper also provides a systematic attribution of the negative result of diffusion-based synthetic augmentation and honest overhead measurements. The takeaway is methodological: class-hierarchy priors offer cheap, deployment-friendly structural regularization, but their claimed cross-domain value must be re-examined on leakage-free evaluation sets.