Advanced Computer Vision Techniques for Detecting and Segmenting Structural Visible Seismic Damages Under Varied Testing Conditions

Abstract

A cutting-edge computer vision method, incorporating an attention mechanism, transformer architecture, and a customized U-Net model, is used for pixel-level multicategory detection of visible seismic damage in reinforced concrete (RC) bridge piers. The damage categories include cracks, spalling, reinforcement exposure, crushing, buckling, and structural failure. A semantic segmentation database is created from experimental photos obtained through cyclic tests, shaking table experiments, and real- time hybrid simulations. To ensure seamless reconstruction, smooth blending techniques such as overlapping and mirror padding are applied to the predicted patch masks. The image database undergoes extensive preprocessing, including lens correction, perspective adjustment, labeling, and damage-type balancing using rotation, flipping, Gamma correction, Hue and Saturation adjustments, and blurring effects. Both sample-level and pixel-level data balancing are achieved through hypergeometric distribution and weighted loss functions, respectively, ensuring the desired probability distribution for each damage category. A hybrid loss function optimizes model performance and metrics like Intersection over Union (IoU) and F1 score track training and validation progress. Atrous convolution is integrated for multi-scale feature extraction, enhancing detection accuracy across varying spatial resolutions. The proposed vision-based approach is validated on unseen (out-of-database) RC bridge pier images, demonstrating high accuracy in detecting multicategory seismic damage. Additionally, crack feature extraction is conducted, measuring total crack length, average and maximum crack widths, and angle, while the location of maximum crack width is also investigated. These findings underscore the promise of automated structural health monitoring and post-earthquake safety assessments, enhancing resilience and enabling rapid decision-making.