PSI - Issue 84

Manuel Mancuso et al. / Procedia Structural Integrity 84 (2026) 1347–1352

1350

Figure 1 Precision vs Recall graph for each tunnel

Notably, crack density is not necessarily correlated with higher IQOA indices: tunnels may present numerous narrow or closed cracks (e.g. shrinkage-related), such as Tunnel #015, may exhibit low noise levels and limited general degradation while remaining challenging for detection due to reduced crack width and then visual contrast. Conversely, tunnels with much higher IQOA values, such as Tunnel #120, often show fewer but wider and well defined cracks, which are more easily recognizable despite elevated visual noise. Overall, these results highlight the combined influence of multiple interacting factors on model performance. By observing confusion matrices across the five case studies, the proportion of correctly detected cracks ranged from 30% to 73%, consistent with the trends observed in the PR curves and reflecting the general influence of visual noise, as a function of degradation, on the learning process. A recurrent limitation across all experiments was the high rate of false-positive (FP) detections: the model frequently predicted the presence of cracks even in images where none were actually present. This behaviour could be related to the objectness-based nature of the YOLO architecture rather than to confidence thresholds or dataset balance. Objectness estimates whether a region appears to contain any object, not specifically a crack. Given that cracks are thin, linear, and often visually similar to the background, texture, edges, or noise could be interpreted as objects. Moreover, YOLO is structurally biased toward high recall, favoring false positives over false negatives, and the background is not sufficiently “hard” to guide the network when cracks are absent This behaviour suggests that the model tends to associate visual noises, such as structural joint, trace of water infiltration, with cracks, highlighting a high sensitivity of the crack detection to background noise in our datasets. 3.2. Semantic Segmentation Two convolutional neural network architectures were investigated for the semantic segmentation of cracks consisting of U-Net model trained from scratch and DeepLabV3 pre-trained network. The objective of the workflow is to obtain a fully segmented crack map for each 50-m tunnel section. In both architectures, crack recognition is