Classical methods for inspecting bridge structures are well established but demanding, both in terms of fieldwork and the processing and evaluation of collected data. The development of artificial intelligence has increased interest in automating this process. In this thesis, using the left abutment of an existing reinforced-concrete highway underpass as a case study (62 segments of 1×1 m), we test the applicability of two open source, pretrained automatic crack-detection models, a classification model, dacl1k (method 1), and a segmentation model, dacl10k, applied in three iterative variants (method 2a, 2b and 2c), and compare their results against a classical, manual inspection. Method 1 showed weak discriminative ability for crack presence (AUC = 0.521), missing approximately 76% of manually confirmed cracks at the conventional decision threshold. For method 2, the bias of the predicted crack length decreased across the three iterations (from -54.7 cm for the baseline variant to 2.2 cm for the final variant, which addresses graffiti-related interference as a colour problem rather than a resolution problem), but paired statistical tests (the Wilcoxon signed-rank test) did not confirm significant differences between the variants, whereas differences in inference time between all variants were statistically significant. The two main causes of the automatic methods' limited reliability were identified as high-contrast graffiti and cracks that had been prepared for repair injection - cases not represented in the models' training data. The results show that the tested methods do not yet match the reliability of classical inspection, but may be useful as a triage tool, directing an inspector's attention toward the segments most likely to be damaged.
|
