Work

YOLOv11-RAH: A recurrent attention-enhanced edge intelligence network for UAV-based power transmission line insulator inspection

Year: 2025

Type: article

Abstract: Highlights A Recurrent Attention Head (RAH) is presented, containing a lightweight depthwise–gated convolution module and an adaptive early-stopping rule. The gated convolution refines features with minimal overhead, while the stopping rule halts iterations once convergence is detected, enabling accurate defect localisation under real-time constraints. In addition, a four stage augmentation pipeline, covering geometric, photometric, Mosaic, and random-crop transformations, is integrated to enlarge data diversity and strengthen model robustness.A consolidated insulator-inspection corpus is organised, combining two cleaned public benchmarks (PTL-AI Furnas-I and CPLID) with a newly curated proprietary dataset (AIDD). The resulting collection delivers diverse outdoor conditions and finely annotated defect categories for comprehensive evaluation.A thorough experimental study is conducted, including cross-model comparisons (SSD, Faster R-CNN, YOLOv5n, YOLOv11 variants) and ablation of the augmentation pipeline. The proposed YOLOv11-RAH attains 81.7% mAP on PTL-AI Furnas-I, 75.7% on CPLID, and 79.0% on AIDD, outperforming the YOLOv11n baseline with no attention modules and data augmentation by about 5 percentage points under identical settings.

Source: International Journal of Intelligent Networks

DOI: 10.1016/j.ijin.2025.11.008