Rail surface defect segmentation via strip-aware deformable convolution and spectral deformable transformer

Abstract

Accurate segmentation of rail surface defect regions is critical to the safe operation of railway systems. To achieve precise rail defect segmentation, a segmentation approach that integrates deformable convolution with a spectral–deformable transformer was proposed. A multi-scale strip-aware orientation-constrained deformable convolution was employed in the upper encoder block, while a Spectral Deformable Transformer (SDT) was introduced in the deepest encoder block. In the decoder network, a feature-gated fusion module was constructed by combining an Efficient Up-Convolution Block (EUCB) with a Mixed Local Channel Attention (MLCA) module. In the training stage, the pixel-wise classification loss was calculated by a label smoothing weighted cross-entropy function, while the object-level loss was calculated by a weighted generalized dice function. On this basis, a unified quality score was introduced to enable coordinated optimization of the two independent loss functions. Based on a self-developed railway inspection platform, rail surface defect images were collected from in-service railway lines and a defect segmentation dataset was constructed. Experimental results on the test dataset show that the proposed method achieved 74.89% mIoU and 85.38% F1. These results demonstrate that the proposed framework can accurately and robustly segment slender rail surface defects under complex inspection conditions, showing practical potential for intelligent railway inspection and maintenance.