Construction and Optimization of Machine Learning Networks for Rail Surface Defect Detection

To ensure safe train operation, the importance of rail management is increasing. To detect rail surface defects and manage rail maintenance accordingly, surface defect images were collected from various railway lines using a portable rail surface defect imaging device. Based on the collected images, three semantic segmentation models were constructed: UNet, PSPNet, and Deeplab V3+. Performance of each model was compared and analyzed by varying database (D/B) sizes and batch sizes. Performance evaluation was conducted using various indicators such as recall, precision, accuracy, IoU, and pixel accuracy to assess defect detection accuracy. Analysis results showed that Deeplab V3+ demonstrated overall superior and stable performance, with optimal performance achieved at a D/B size of 116 images or more and a batch size of 4 or greater. The findings of this study can serve as fundamental data for constructing an efficient rail surface defect detection system.