Complementary modules and the Precision–Recall trade-off in pavement crack segmentation examined through a controlled decomposition of FPN-CAS

Pavement crack segmentation is constrained by a Precision-Recall trade-off that single-axis architectural changes have not resolved. We test whether modules with opposing trade-off vectors integrate complementarily. We built FPN-CAS, a ResNet-34 Feature Pyramid Network combining ASPP (context), CBAM (attention) and Dynamic SnakeConv (geometry), and decomposed it into seventeen controlled configurations, each trained with three seeds. A merged pool of 1,413 images (Crack500, CFD, AsphaltCrack300) was split at image level into 1,130 training, 141 evaluation and 142 reserve images. On the evaluation subset, FPN-CAS raised IoU from 63.94 ± 0.16% to 66.07 ± 0.43%. SnakeConv alone lost 1.00 pp of Recall, yet in the integrated model this became a gain. IoU, Dice and Recall gains were super-additive, and removing any module cost at least 0.91 pp of IoU. However, the gains were fragile. On the reserve subset, FPN-CAS lowered IoU by 1.07 ± 0.54 pp and Precision by 3.58 ± 1.31 pp in all three seeds. Source-wise IoU change ranged from +8.52 pp (AsphaltCrack300) to −17.86 pp (CFD), and no per-image difference survived multiple-comparison correction. Rather than proposing a new module, we offer a controlled decomposition that separates complementary from redundant modules under one protocol. The observed complementarity is confined to the evaluation subset and does not establish dataset-independent superiority.