Detection of High-Speed Railway Track Bed Foreign Object Based on Contrastive Learning and Adversarial Training

High-speed railway safety is compromised by track bed foreign objects, highlighting the imperative for efficient inspection. The track bed foreign objects detection is impeded by the rarity and diversity of those objects, resulting in challenges posed by limited and uncharacterized sample sets. This paper proposes a novel detection method based on contrastive learning and adversarial training, preserving the detection speed of one-stage detection while substantially enhancing accuracy. Contrastive learning sharpens feature discrimination, and adversarial training generates robust samples to bolster model performance. Experiments demonstrate that this approach yields a 12.9% increase in mAP.5 and 9.2% increase in mAP.5:.95, all without additional model parameters.