From Digital Model to Reality Application: A Domain Adaptation Method for Rail Defect Detection

Wenkai Cui, Jianzhu Wang, Haomin Yu, Wenjuan Peng, Le Wang, Shengchun Wang, Peng Dai, Qingyong Li

Abstract: Recently, vision-based rail defect detection has attracted much attention owing to its practical significance. However, it still faces some challenges, such as high false alarm rate and poor feature robustness. With the development of deep neural networks (DNNs), deep learning based models have shown the potential to solve the problems. Nevertheless, these models usually require a large number of training samples, while collecting and labeling sufficient defective rail images is somewhat impractical. On the one hand, the probability of defect occurrence is low. On the other hand, we are not able to annotate samples that include all types of defects. To this end, we propose to generate defective training images in the digital space. In order to bridge the gap between virtual and real defective samples, this paper presents a domain adaptation based model for rail defect detection. The proposed method is evaluated on a real-world dataset. Experimental results show that our proposed method is superior to five established baselines.