Semi-Supervised YOLOv8 for Airport Runway Foreign Object Detection

Foreign object detection (FOD) on airport runways is critical for aviation safety, but small, low-contrast objects remain challenging for traditional supervised methods due to limited labeled data and annotation biases. This paper proposes a semi-supervised framework combining MoCo v2 self-supervised pretraining, dynamic pseudo-labeling, and SAM v2-based augmentation to reduce reliance on labeled data while improving small object detection. Experiments show our method achieves a 73.7% mAP and 56.6% SDR, outperforming state-of-the-art approaches by 8.6% mAP. The results validate the effectiveness of leveraging unlabeled data and adaptive augmentation in complex runway environments.