Synthetic‐to‐realistic domain adaptation for cold‐start of rail inspection systems

Abstract Rail surface defects are potential danger factors for railway systems, and visual inspection of plays a vital role in rail maintenance. Recently, the methods based on deep learning have been widely used but such systems often face problem lack defect samples training models at start‐up, which is called cold‐start problem. It challenging to obtain sufficient since sparse even non‐existent running system. Therefore, synthetic‐to‐realistic domain adaptation (SRDA) method proposed real‐world inspection. SRDA adapts synthetic images look more realistic reducing gap between obtains translated that consist information background. After that, train detector defects. In order make robust complex backgrounds, generates with same defect‐level semantics different texture appearances makes align these learned feature space. addition, (SRRSD) dataset containing 20,662 built. The experimental results SRRSD show achieves higher detection performance than other established adaption 19.0% 26.7% average precision.