CCVS-3ASeg: a large-scale contextual auxiliary framework for road crack semantic segmentation in UAV imagery

Existing road crack inspection using UAVs and vehicle-mounted systems often relies on local field-of-view observations, requiring repeated scanning in large road scenes. We present CCVS-3ASeg, a framework for a preliminary feasibility evaluation of UAV-based large-scale road crack semantic segmentation. Its design focuses on two coupled challenges: computational overhead from large-scale feature representation and loss of fine crack features during spatial compression. CCVS extracts road regions and performs direction-aligned slicing, while 3ASeg uses an Ultra-Lightweight Adversarial Attention Mechanism (Ultra-LAAM) to assist crack-feature representation. Evaluation comprised CCVS road processing on the Custom-Dataset, 3ASeg-assisted segmentation on Crack500 and a selected public-source Crack-Dataset, and a two-image G94 Expressway case study of the connected workflow. Across test images, CCVS achieved recall of 0.9377 ± 0.1517 and pixel accuracy (PA) of 0.9606 ± 0.1010. On Crack500, the largest observed mIoU gain among evaluated 3ASeg-assisted configurations was 6.93 percentage points, and Ultra-LAAM reached 1132.27 ± 18.15 FPS. These component‑level experiments preliminarily demonstrate that large‑scale road extraction to assist pavement crack segmentation is a viable direction for further exploration and optimization.