Railway Structure Diagnosis Based on Swin-Transformer Backbone with Mask R-CNN

Abstract

As the fast development of city metro construction, railway system safety has become one of the most popular discussions in metro engineering. The components of railway structure are likely to be air-slaked or breached while the metro system is running. Till now the essential components maintaining is mostly done by human, and this costs time and human resources. CNN has high accuracy on object detection but may not be able to locate the minor damage happened on railway structure components. In this study, a Swin Transformer based method is introduced for components diagnosis in railway system, which can detect and segment the target object, even they are sometimes small and indistinguishable for human eyes. The model with best performance reached the bbox mAP of 0.546 and segm mAP of 0.443, and it can be deployed on a laptop with camera.