Automated rail surface crack analytics using deep data-driven models and transfer learning

Abstract

A deep transfer learning (DTL) framework for the rail surface crack detection using a limited volume of training images is developed in this study. In DTL, two deep learning models pre-trained by two recently popular object detection algorithms, YOLOv3 and RetinaNet, based on the COCO dataset are considered as base models for next transfer learning via using rail track images. Transferred YOLOv3 and RetinaNet models are separately applied to detect rail track cracks. Based on their outputs, an ensemble scheme is developed to further improve the detection performance. The effectiveness of the model developed by the DTL is validated by benchmarking against models developed by YOLOv3, RetinaNet, faster R-CNN (F-RCNN) and single shot multibox detector (SSD) separately as well as two conventional rail surface crack detection methods, the visual detection system (VDS) and the geometrical approach (GEA). A dataset composed of 35 rail images from the China Railway Corporation and 67 rail images from the publicly available Type-I RSDDs dataset is considered. Computational results show that the model developed by DTL attains the best recall and average precision comparing with benchmarking models. VDS and GEA almost fail in the crack detection task due to the complex background and various noises in images. In addition, in terms of the average precision, YOLOv3 model is more suitable for detecting cracks of small sizes while RetinaNet model performs better on detecting cracks of large sizes.