PSI - Issue 62

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

135

7

= = + = 1 ∑ = 1 +

(1)

(2)

(3)

Looking at the obtained results, both models can predict defects, even if some differences can be appreciated. To difference of the work provided in Cardellicchio et al. (2023b), no different conditions were used to achieve higher performances (e.g., higher number of pixels). Also analyzing the confusion matrices of the models, here not reported, it is worth mentioning their diagonality, which implies that defects are correctly recognized with different accuracy, with high performance provided by YOLOv5x. Observing each class (i.e., defects), the graphs in Fig. 2 show that for some defects the obtained precision is different between the two showed models, as occurs for pavement degradation and shrinkage cracks. In general, YOLOv5x achieves higher precisions with low confidence than YOLOv5n. Observing the validation test reported in Fig. 3, there is a good number of missed detections, especially for the defects with lower labels (i.e., pavement degradation and shrinkage cracks). Instead, very few differences exist in terms of defect detection within the same image, which means that both models are trained in order to predict the same bounding box for the identified class. Overall, from the qualitative point of view, most of the bounding boxes identifying defects are correctly localized, with exception of some cases, such as in the first three figures of the collage presented (i.e., tags 4432, 4497, 4589). In the end, although the different metrics, both models can localize most of the defects with a relatively high confidence score. 4. Conclusions and future developments The work presents a study on the use of single-stage detector for predicting the most typical defects in RC existing bridges, having at disposal images. In detail, a dataset of figures reporting typical defects in existing RC bridges was collected, and authors as domain experts performed a labelling according to few classes, i.e., cracks, corroded and oxidized steel reinforcements, deteriorated concrete, honeycombs, moisture spots, pavement degradation, and shrinkage cracks. Thus, several versions of the most fairly popular single-stage detector, YOLO, were trained on the dataset, by using the architectures available for the version 5. Results were evaluated in terms of usual metrics, such as precision, recall and mean average precision, for all classes. From the experimental campaign, YOLOv5x presented the higher performance, while YOLOv5n presented the lower performance. Comparisons between the best and the worst networks was provided, in terms of precision-confidence (for each class and in average) and by looking the capacity of prediction on some images of the database. The obtained results show that both networks are able to detect damages in a photo, although some differences exist. In addition, the imbalance of some classes, such as pavement degradation and shrinkage, does not allow to yet achieve a good prediction capacity for all considered defects. Further developments will be aimed at several main goals: (a) to enlarge the available dataset, in order to provide more data that can be more balanced than the currently available ones; (b) to use more denser networks for achieving higher values of precision and recall; (c) to test the most recent versions of the YOLO (e.g., version 8); (d) to propose a tool based on YOLO, to be used for periodical inspections, which can be automatically processed, as for example by employing unnamed aerial vehicles. Acknowledgements All authors thank the consortium FABRE for the financial support. The first author acknowledges funding by Italian Ministry of University and Research, within the project ‘PON -Ricerca e Innovazione 2014 – 2020, (D.M. 10/08/2021, n. 1062) CUP CODE: D95F21002140006.