Railway Track Multicomponent Segmentation Based on Residual Contextual Transformer

Regular inspection of track component conditions is crucial for maintaining railway safety, and accurate segmentation of track multicomponent is fundamental to the detection and severity evaluation of track diseases. The U-shaped architecture has achieved tremendous success in segmentation tasks. However, most existing U-shaped architecture-based methods are not suitable for tracking multicomponent segmentation due to the limitations in explicitly modeling long-range dependency. In order to address this problem, this article proposes an innovative track multicomponent segmentation (TMCSeg) network based on the residual contextual transformer (Res-CoT) and adopts the encoder-decoder framework. In the encoder stage, the long-range rich context information can be obtained by adopting designed Res-CoT blocks and cascading transformer layers. In the decoder stage, a multiscale feature fusion (MFF) module is used to aggregate the decoding features of different scales to predict segmentation results. Especially, the decoding features of each scale are obtained by integrating the coding features of the corresponding scale. Numerous tests have been carried out on constructed track datasets. The experimental results demonstrate that our proposed network exhibits strong segmentation capabilities for tracking multicomponent and surpasses the benchmarks in metrics of mIoU and mDice.