Semantic Segmentation of Bridge Cracks Using an Improved SegFormer Architecture

As a critical component of railway infrastructure, the structural safety of railway bridges plays a vital role in ensuring the overall safety of rail transport. To address the problem of crack detection in bridges, this paper presents a crack segmentation framework for railway bridge inspection, based on an enhanced SegFormer architecture. To tackle the challenges of non-linear crack patterns, weak feature saliency, and scale variation, we propose three key modules: Dynamic Snake Convolution (DySnake) for geometric modeling of winding cracks, a dual-domain attention mechanism (MLCA) for enhancing low-contrast features, and an Attentional Feature Fusion (AttFF) module for multi-scale feature integration. The entire framework is built on a Transformer-based encoder-decoder structure, enabling efficient semantic segmentation in complex backgrounds. Experiments on the CrackSeg3 dataset show that the proposed method achieves a mean IoU of 78.98%, outperforming existing baselines. Ablation studies demonstrate that DySnake and AttFF individually contribute 0.92%–3.67% mIoU improvement. The results confirm that the model achieves accurate, fine-grained crack detection through pixel-level contour localization, making it a reliable tool for automated structural health monitoring.