PSI - Issue 84

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

1349

3.1. Object Detection (OD)

Five tunnels with heterogeneous structural and environmental conditions were analysed to evaluate OD performance (Tab. 1). Orthophotos were subdivided into fixed-size tiles, and crack polylines were converted into bounding boxes for YOLOv5 fine-tuning. Hyperparameters were progressively optimised through tile sizing, data augmentation, and increased dropout to improve convergence and reduce overfitting (Tab. 2).

Table 1. Tunnel Metadata. Year of construction/renovation

Tunnel ID Label index Quality

Noise index Quality

Impermeabilization N° images / fine tuning

*210 *120 *125 *015 *220

2 2 2 3 1

4 2 3 1 1

1967 1973 1973

NO NO NO

2410 5758 1825 8006 4314

1965/2000

YES YES

1993

Table 2. Dataset configuration for YOLOv5 fine-tuning

Dataset Configuration Image source Resolution

Orthophoto laserscanner 3-4 mm/pixel Greyscale RGB (0-255)

Type

Epochs

200 500

Imgsz

Batch Conf Dropout Data augmentation -Stride (pre-processing phase)

12 0.25 0.15 50%

Model performance was evaluated using precision–recall (PR) curves computed for each tunnel. Although the performance varies depending on the specific tunnel, the PR trends show a broadly consistent behaviour across datasets (Fig. 1). Detection capability generally decreases in tunnels characterised by higher level of noise, mainly associated with hydraulic and structural degradation, as well as with a lower label quality. In addition, performance tends to be influenced by the number of images available for each tunnel, which depends not only on tunnel length but also on crack density, as the dataset was balanced between crack-containing and non-crack image tiles.