Robust crack detection in complex slab track scenarios using STC-YOLO and synthetic data with highly simulated modeling

Crack detection in slab tracks plays a crucial role in accident prevention. Existing algorithms primarily operate on monotonous concrete backgrounds and often struggle with data scarcity and complex scenes. This paper proposes a parametric slab track model replicating real-world inspection conditions through high-fidelity virtual simulation, enabling realistic synthetic crack data generation. The subsequently developed STC-YOLO network utilizes these synthetic images to enhance fine crack detection in complex slab track scenes. Results show that STC-YOLO trained on synthetic data (4:1 virtual-to-real ratio) achieves over 20 % improvements in both mAP and recall compared to using no virtual images, outperforming traditional augmentation methods like horizontal flipping and color dithering. Moreover, STC-YOLO exhibits over 6 % higher mAP than the baseline algorithm and surpasses five state-of-the-art object detection networks. The proposed algorithm greatly reduces the cost of data acquisition. • Highly simulated virtual model built to reproduce real-world inspection behavior. • A crack image synthesis approach across complex scenes is designed. • STC-YOLO, an object detection network for slab track cracks is proposed. • Contribution of introduced virtual data for deployment of deep learning is demonstrated.