Utilizing Transformer-Based Semantic Segmentation for Damage and Component Detection in Civil Infrastructure Inspection

Semantic segmentation plays an essential part in the condition assessment of bridges by enabling pixel-level classification of structural components and defects. Compared with U-shaped models, transformers excel at capturing image context by leveraging self-attention and parallel processing, leading to superior performance. In this research, a novel approach was introduced by integrating the SegFormer (MiT-B5) model into the railway bridge inspection task. The study utilized the Tokaido dataset to train and test the SegFormer (MiT-B5) model for component and damage detection tasks. Experimental results reveal that SegFormer considerably outperforms conventional U-Net and Attention U-Net baselines, obtaining a mean intersection of union (mIoU) of 72.89%, compared with 71.50% (U-Net) and 64.77% (Attention U-Net) in component detection tasks. Furthermore, the SegFormer model obtains an mIoU of 60.22%, compared with 43.51% (U-Net) and 45.92% (Attention U-Net) in the damage detection task. The model exhibited considerable improvements in the segmentation of structural classes such as slabs, beams, and columns. The study provides a detailed class-wise performance analysis, emphasizing challenges in detecting minority classes and the need for improved data representation. This research advances the use of transformer models in structural health monitoring and also paves the way for scalable, unmanned aerial vehicle–compatible inspection systems.