RTCSeg: A Novel CoT-Trans-Based Multi-Class Segmentation Network for Railway Track Components

Railway track lines are mainly composed of fasteners, sleepers, and rails. The safety status of each component affects the safety of train operation. Therefore, for subsequent state monitoring and defect detection, it is important to simultaneously segment multiple components of railway tracks. In order to address the problem of accurately segmenting multiple targets simultaneously, this paper proposes an innovative railway track component multi-class segmentation network based on cot trans (referred to as RTC). The network is an encoder-decoder network with a TRANS structure, and the cot block in the encoder not only has the characteristics of self-attention mechanism but also captures adjacent contextual information, enabling the encoder to extract more effective features. The decoder incorporates a multi-scale feature fusion module, which can fuse deep and shallow features, resulting in more accurate segmentation results. Meanwhile, two datasets containing images with fasteners, sleepers, and rails were constructed in this study. Through quantitative and qualitative experiments, it is demonstrated that the proposed RTCSeg network can accurately segment multiple components of railway tracks in images, outperforming other compared methods. Additionally, ablation experiments further validate the effectiveness of the proposed module.