AE-YOLO detects insulator defects with 95% mAP using UAV imagery
New framework combines autoencoders and YOLO for 5% mAP improvement over baselines
Deep Dive
Proposed by Malak Allam, Khaled Shaban, and Ali Hamdi, AE-YOLO is an attention-guided autoencoder-enhanced YOLO framework for detecting small insulator defects in UAV transmission-line images. It integrates lightweight bottleneck autoencoders within an FPN-PAN neck and CBAM attention modules. Using an EfficientNetV2 backbone, it achieves 95.10% mAP@0.5, 96.40% precision, and
Key Points
- Achieves 95.10% mAP@0.5 and 93.80% recall using EfficientNetV2 backbone on insulator defect dataset
- Combines predictions from YOLOv8, YOLOv10, and YOLO11 via Weighted Boxes Fusion with autoencoder confidence boosting
- Surpasses best YOLO baseline by 5.0 mAP points and 6.7 recall points, addressing class imbalance and small defect detection
Why It Matters
Enables reliable drone-based detection of tiny power line defects, reducing manual inspection costs and improving grid safety.