Pixel-level multicategory semantic segmentation of visible seismic damage in bridge piers using an Attention-Mamba Transformer-based U-Net model

Abstract

Currently, many computer vision-based studies focus on cyclic test photos approximating structural behavior under seismic loads and struggling with severely imbalanced multiclass seismic damage detection, particularly cracks. This paper presents an approach for pixel-level detection of visible seismic damage in RC bridge piers, identifying cracks, spalling, reinforcement exposure, crushing, and buckling/failure. A semantic segmentation database is built from experimental images emphasizing real-time hybrid simulations, with lens correction, perspective adjustment, and augmentation to enhance diversity. Hypergeometric distribution and weighted loss functions address class imbalance at both sample and pixel levels. A self-attention-Mamba-driven transformer block with inception modules is integrated into a customized U-Net bottleneck, achieving per-class IoU over 0.7958. A VGG16 encoder with Mamba blocks further refines crack feature extraction (length, width, angles), reaching IoU above 0.6478. Overlapping and mirror padding improve mask blending. The model generalizes well to unseen bridge piers and shear walls, supporting accurate post-earthquake damage assessment.