An efficient multi-cue feature fusion framework and benchmark for robust UAV-based insulator defect detection

Timely detection of insulator defects is critical for reliable power system operation; however, UAV-based inspection in real-world environments is often hindered by limited benchmark coverage, dynamic environmental interference, inadequate small-target representation, and localization inaccuracies. To facilitate practical algorithm development and evaluation, we establish the I nsulator D efect D etection mi X ture ( IDDX ) benchmark, which comprises 16,540 UAV images covering three common defect categories under diverse real-world conditions. Building upon IDDX, we develop a YOLO-based detection framework, YOLO-IDDX , incorporating two dedicated designs for robust and efficient deployment. First, the proposed M ulti-branch I nteractive e X traction Down sampling ( MIXDown ) module jointly captures complementary semantic, edge, and frequency cues, thereby enhancing small-target feature representation under complex backgrounds. Second, we propose a T ightness-constrained I ntersection o ver U nion ( TIoU ) loss to suppress oversized predictions, thereby improving localization compactness and stability while reducing false positives. When integrated into YOLOv8, the nano variant of YOLO-IDDX achieves 90.2% mAP50, 66.9% mAP75, and 59.3% mAP50:95 on IDDX, outperforming YOLOv8n by +1.8/+2.1/+1.6 percentage points while using only 89% of its parameters. Extensive ablation studies further validate the effectiveness of MIXDown and TIoU in challenging scenarios, indicating that the proposed benchmark and framework offer a practical solution for intelligent UAV-based insulator inspection.