ABSTRACT Railroad safety is paramount for the seamless operation of rail transit systems. With the development of artificial intelligence, numerous convolutional neural network (CNN) models have been created and applied to the detection of railroad surface defects. However, the limited edge computing power of embedded devices poses a significant challenge to the effective deployment of most models in practical applications. In this study, we demonstrate that the defect segmentation task relies heavily on shallow feature extraction and propose an exceedingly lightweight semantic segmentation model, RailNet, specifically tailored for the detection of railroad surface defects. RailNet boasts a parameter count of merely 0.905 M and occupies only 2.45 MB when deployed on an embedded development device. This model achieves superior segmentation performance compared to several classical semantic segmentation models. Furthermore, to address the issues of poor image quality of track surface defects and the challenges of dataset collection in real‐world scenarios, this study introduces the chaotic differential background modelling (CBDM) image quality enhancement algorithm and a GAN‐based dataset enhancement method. The proposed semantic segmentation network and dataset enhancement approach are validated on the RSDDs (rail surface defect detection) dataset, achieving an optimal balance between segmentation performance and computational cost.
Wu et al. (Thu,) studied this question.