Self-Supervised Defect Representation Learning for Label-Limited Rail Surface Defect Detection

Abstract

An automatic detection method for surface defects on railway tracks holds significant importance in ensuring the safety of railway transportation. However, in practice, defects on railway tracks exhibit characteristics, such as being scarce in number, small in size, and having significant shape variations. Therefore, implementing supervised learning techniques under the constraint of limited labeled data is a major challenge. To address this problem, we propose a designed framework based on self-supervised representation learning for rail surface defect detection (R-SSRL). Inspired by deep neural networks, the R-SSRL is organized based on a convolutional encoder–decoder neural network to segment rail defects. Also, it uses a novel self-supervised algorithm and a designed defect simulation method to learn possible feature representations of defects from defect-free rail samples. This enables the R-SSRL to utilize defect-free samples that are readily available, to improve model performance with limited labeled data. Experiments on a real-world dataset show that the R-SSRL framework exhibits superior performance in the rail defect detection task, outperforming other models.