Railway Track Crack Identification via Stochastic Data Augmentation and YOLO Framework

Effective railway track inspection is essential formaintaining train safety and operational reliability. Critical safetyindicators include track cracking, excessive vegetation growth,gauge separation, and track surface defects. In the given scenario,the publicly available image dataset is limited in size. Generatingan augmented dataset that effectively mitigates the risk ofoverfitting remains a critical challenge. In this paper, we proposea novel data augmentation approach for railway track crackdetection, leveraging stochastic modeling based on establishedpublicly available image transformation techniques. It is alsoassumed that the drone/UAV transmits real-time track images,and both image processing and the transmission of processedinformation to the end user occur in real time. The other aspect ofthe work focuses on the accurate identification of the target object(railway track cracks) in the captured imagery using appropriatemachine learning techniques. Accordingly, the YOLO frameworkhas been extensively explored, and the simulation results validateits accuracy and computational efficiency.