Pyramid–MixNet: Integrate Attention into Encoder-Decoder Transformer Framework for Automatic Railway Surface Damage Segmentation

Rail surface damage is a critical component of high-speed railway infrastructure, directly affecting train operational stability and safety. Existing methods face limitations in accuracy and speed for small-sample, multi-category, and multi-scale target segmentation tasks. To address these challenges, this paper proposes Pyramid-MixNet, an intelligent segmentation model for high-speed rail surface damage, leveraging dataset construction and expansion alongside a feature pyramid-based encoder-decoder network with multi-attention mechanisms. The encoding network integrates Spatial Reduction Masked Multi-Head Attention (SRMMHA) to enhance global feature extraction while reducing trainable parameters. The decoding network incorporates Mix-Attention (MA), enabling multi-scale structural understanding and cross-scale token group correlation learning. Experimental results demonstrate that the proposed method achieves 62.17% average segmentation accuracy, 80.28% Damage Dice Coefficient, and 56.83 FPS, meeting real-time detection requirements. The model’s high accuracy and scene adaptability significantly improve the detection of small-scale and complex multi-scale rail damage, offering practical value for real-time monitoring in high-speed railway maintenance systems.