RTCSeg: An End-to-End Multi-Class Segmentation Network for Railway Track Components

First, this paper constructs a dataset of railway track components (dataset 1 and dataset 2), and then proposes an end-to-end railway track component multi-class segmentation network called RTCSeg, which is trained and tested on the constructed dataset. The proposed RTCSeg network consists of an encoder and a decoder, where the encoder incorporates Contextual Transformer blocks and Transformer blocks to extract global and local feature information from the input image. The decoder introduces a multi-scale feature fusion module to merge multi-layer information and obtain richer feature information, thereby improving segmentation performance. Extensive experiments based on the created dataset demonstrate the effectiveness of the proposed CoT block and multi-scale feature fusion module. Moreover, the proposed RTCSeg can effectively and accurately segment multi-class images of railway track components, providing a good method and approach for practical applications in railway segmentation and detection, and improving the intelligence level of multi-object segmentation and subsequent defect detection in track components.
In the future, we plan to expand the dataset of railway track component images. Additionally, the inference speed of the proposed segmentation network in this paper is slow, and we will optimize the model to make it faster in the future.