A Dual-Stage Deep Learning Framework for Fine-Crack Segmentation in High-Resolution UAV Images for Bridge Inspection

Fine-crack segmentation in high-resolution UAV images of bridges is hindered by weak crack signals, complex backgrounds, and substantial computational demands. This study proposes a dual-stage framework combining lightweight grid-wise screening with graph-enhanced segmentation. CrackClsNet identifies candidate regions, and CrackSegNet processes contextual patches comprising each candidate and its eight neighbors. Multiscale and deformable convolutions, attention, and cross-patch graph reasoning support feature extraction, while overlapping predictions are fused during reconstruction. In experiments using 1000 pixel-annotated crack-containing images collected from 13 highway bridges and 500 crack-free images, neighborhood retention increased candidate region crack-pixel coverage from 91.97% to 95.65% on the classification test set. The end-to-end framework achieved an F1-score of 0.5826, Crack IoU of 0.4110, and mIoU of 0.7052, exceeding six baseline methods in these metrics. The results of ablation experiments further support the benefits of neighboring context and cross-patch feature exchange, with 3 × 3 neighborhoods balancing segmentation quality and computational cost. With approximately 4.41 million parameters, the framework achieved 4.10 FPS with a reported memory consumption of 6.94 GB at 4096 × 5632 pixels, providing the highest throughput and lowest reported memory consumption among the evaluated methods at this resolution. These results indicate a promising balance between segmentation performance and computational cost on the evaluated bridge image dataset.