Answers are grounded in this paper and cite their sources.
Transmission-line component and defect detection remains challenging because of variations in object scale, complex backgrounds, class imbalance, and image perturbations. This study proposes FA-YOLO, a lightweight detector based on YOLO11n. Its FADConv module combines FANet-based channel and spatial feature recalibration with AKConv-based adaptive sampling, while a pyramid-pooling squeeze-and-excitation (PPSE) module integrates multi-scale contextual information and recalibrates channel responses. The training objective combines Focal Loss and WIoUv3 for classification and bounding-box regression, respectively, while retaining Distribution Focal Loss. FA-YOLO was evaluated on a private dataset containing 10,561 images across 18 categories and on the public Chinese Power Line Insulator Dataset (CPLID), with comparisons against general-purpose detectors and two insulator-specific models, IDD-YOLO and YOLOv8-eRFD-AP. On the private dataset, FA-YOLO achieved mAP@0.5 values of 82.4%, 70.8%, and 65.7% on clean images, Gaussian-noise-degraded images, and adversarial images generated using a 20-step white-box PGD attack with an L∞ perturbation bound of 8/255, respectively. These results exceeded those of YOLO11n by 3.6, 3.0, and 5.4 percentage points. Across three training seeds, clean-image mAP@0.5 averaged 81.57% with a sample standard deviation of 0.85 percentage points. Class-specific evaluation also showed improvements in three underrepresented categories, although substantial missed-detection rates remained for scarce defects. On CPLID, where all compared models were retrained, FA-YOLO achieved an overall mAP@0.5 of 93.5% and mAP@0.5:0.95 of 62.4%, exceeding YOLO11n by 0.9 and 2.9 percentage points. These findings support the model’s applicability to both datasets after training and improved detection performance under the evaluated perturbations, without establishing robustness beyond the tested attack settings.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: