Abstract
Cracks on concrete and asphalt surfaces are an early warning that a dam, a bridge, or a road is starting to fail, and missing them can be expensive. Inspection is increasingly done with drones, which produce far more images than an engineer can check by hand, so the analysis has to be automated. In this paper we compare five semantic segmentation networks for crack detection and, unlike most such comparisons, we push all of them through exactly the same pipeline. Four are convolutional (U-Net, UNet++, DeepLabV3+, and a Feature Pyramid Network) and one is a transformer (SegFormer-B2). Every model sees the same data, the same 320 x 320 inputs, the same augmentation, the same Binary Cross-Entropy plus Dice loss, and the same 0.5 threshold, and all are trained with mixed precision on the public Crack500 benchmark and its official train, validation, and test split. We repeat each training five times with different seeds, report the mean and standard deviation on the test set, and check the gap between the two best models with a paired significance test. To see whether the ranking survives outside Crack500, we then run the trained networks, untouched, on a second dataset (DeepCrack). SegFormer-B2 was the most accurate model on Crack500 (IoU 0.595, F1 0.746), and the t-test shows its margin over UNet++ is well beyond chance. It stayed first on DeepCrack too (IoU 0.704, F1 0.826). We also report parameters, FLOPs, inference time, and memory, so the accuracy can be weighed against the cost of running each model.