PSI - Issue 84

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

1351

formulated as a binary pixel-wise classification problem, where each pixel is assigned either to the crack or background class. The orthophotos were first subdivided into 512 × 512-pixel tiles, which were used as segmentation model inputs. Following segmented images output were re-assembled into a continuous crack mask for the full tunnel section. This was achieved through a SUM-based algorithm, which integrates overlapping predictions to improve spatial continuity. A subsequent post-processing stage was applied to refine the crack geometry, including smoothing operations and clustering-based filtering to remove residual noise, specifically small false-positive clusters. Initial experiments and tests on segmentation models revealed that, like the object-detection task, false positives represented the main source of misinterpretation of the model. To mitigate this issue, we adopted different strategies to improve the segmentation as is shown in Figure 2. A 50% stride was introduced during the tiling stage, allowing each region of the tunnel lining to be analysed multiple times from different spatial contexts reducing edge-related issue and improving crack continuity. Then the loss function was changed from Binary Cross-Entropy to Dice loss, which proved more effective for highly imbalanced datasets in terms of number of pixels representing crack with the respect to the total image area. Finally, dataset re-balancing was applied to reduce bias towards the background class.

Fig. 2. Example of tunnel cross section segmented – U-net vs DeepLab3 for false positive mitigation. The analysis shown in this paragraph were conducted in detail on the tunnel *220 that had previously shown the most favourable conditions by observing metadata and object detection results that offer a global contextual understanding across all over the tunnels. Among the tested models, DeepLabV3 provided the most stable behaviour, with fewer false negatives and improved geometric continuity of cracks, while U-Net particularly benefitted from stride, loss and re-balancing optimisation. In the best-performing case, the combined strategy yielded approximately 74% correctly classified pixels, with false positives reduced to about 18%, indicating encouraging prospects for large-scale deployment. 4. Discussion and conclusions This study compared two deep learning approaches for computer vision with the aim of identifying the most suitable framework for crack detection on tunnel linings using large-scale acquisition systems. Analyses were conducted on a dataset derived from laser-scanner–based grayscale orthophotos, with the goal of building the base for developing a CNN-based prototype that is both technically reliable and operationally scalable. While object detection