PSI - Issue 62

Sergio Ruggieri et al. / Procedia Structural Integrity 62 (2024) 129–136 Author name / Structural Integrity Procedia 00 (2022) 000 – 000

132

4

recognized within an image instead of scanning the entire image, and then, by strongly reducing the related computational effort. The most popular approaches are represented by two- and single-stage detector approaches. The first one is represented by the R-CNN (Girshick et al. 2014), while the second one is represented by YOLO (Redmond et al, 2016). Kim et al (2018) employed transfer learning to train the original R-CNN network for a dataset of 384 images related to cracks in RC bridges, for purpose of qualitative assessment. Still, Cha et al. (2018) used a faster R CNN for detecting five types of cracks and authors trained the model on a dataset of just over 2000 images, achieving high precision. Very few works were developed by using YOLO. An example is provided by Maeda et al. (2018), which employed different single stage detectors with a dataset made by about 9000 images with 15000 instances of road surface damages. Finally, in Ruggieri et al. (2023) and Cardellicchio et al. (2023b), the authors tested single-stage detector for the same dataset proposed in Cardellicchio et al. (2023a) and to predict the same defects. On this base, in the following, the present work reports the achievement obtained by employing different architectures of YOLOv5. 3. Use of YOLOv5 on a dataset of RC bridge defects To apply the proposed methodology, a dataset of 6580 images was collected, which contains defects in the structural elements of existing Italian bridges (e.g., desk and beams, pillars, supports, abutments). Authors of the paper performed labelling by identifying the most typical defects in the images and then, by avoiding to apply the high discretization proposed by the Italian guidelines (MIT, 2021). In fact, defects were grouped for macro-categories and values of intensity were not assigned. As an example, Italian guidelines suggest to identify cracks (i.e., vertical, horizontal diagonal), but for the case at hand only the defect “ cracks ” was considered. In the labellin g, more defects could be found in the same image, which led to have about 10831 instances. Seven defects were considered, among which cracks, corroded and oxidized steel reinforcements, deteriorated concrete, honeycombs, moisture spots, pavement degradation, and shrinkage cracks. In Fig. 1 some photos of the dataset are shown, for which labelling was processed and specific defects were recognized. Instead, Table 1 shows the number of labels for each considered defect.

Fig. 1. Example of images in the dataset, on which labelling of defects was performed. From left to bottom, corroded and oxidized steel reinforcements, moisture spots, and cracks.

Table 1. Number of labels for each defect for the available dataset. Defect

No. of labels

Cracks

1138 2928 2448 1165 2962

Corroded and oxidized steel reinforcement

Deteriorated concrete

Honeycombs Moisture spots

Pavement degradation

119

Shrinkage cracks

71