This study proposes an end-to-end deep learning framework that markedly improves the accuracy, efficiency, and field applicability of unmanned aerial vehicle (UAV)-enabled bridge deck inspections. By combining deck localization with pixel-level crack detection and quantification, the framework provides a unified solution that overcomes the limitations of existing manual and semiautomated inspection methods. The models are deployed using a standalone software package, enabling on-site implementation without programming expertise and supporting practical integration into structural health monitoring workflows. The methodology consists of two primary stages. In the first stage, bridge deck localization is performed using semantic segmentation with a DeepLab v3+ architecture and a ResNet-18 backbone, enhanced through transfer learning on high-resolution UAV imagery. This model achieves a global pixel accuracy of 98.63% and an intersection over union of 97.18%, ensuring robust deck identification across diverse environmental conditions. In the second stage, crack detection and quantification are conducted using a U-Net-based segmentation pipeline. A dual-phase automated labeling strategy is introduced to streamline large-scale data set generation, allowing the U-Net model to attain a Dice similarity coefficient of 85.32% and a pixel-level accuracy of 95.04%. By integrating high-precision deck segmentation, scalable labeling, and detailed crack analysis into a deployable software platform, this framework establishes a robust tool for UAV-based bridge inspections. The approach enhances the reliability and consistency of condition assessments and provides a foundation for future integration with predictive maintenance systems and digital twin models for infrastructure management.
Almasi et al. (Wed,) studied this question.