PSI - Issue 66

Andrii Kompanets et al. / Procedia Structural Integrity 66 (2024) 388–395 Author name / Structural Integrity Procedia 00 (2025) 000–000

390

3

annotation errors impact a neural network performance in crack segmentation was presented in (Xu, 2024). The study found that even a 20% annotation error rate could lead to a reduction of only 6% in the F1 -score. A distinctive feature of this dataset as compared with many other datasets of images of cracks is the complex background of the images. In steel bridges, cracks usually occur near a stress concentrator such as a welded joint, bolt hole, curvature region, etc. Often some features of the bridge structure may appear as a crack on the images. Moreover, on many images physical inspector markings are presented, indicating crack tips, their size, and inspection dates. Additionally, because these images were taken from structures in service, they often contain visual artifacts such as dust, dirt, raindrop spills, insects, and spider webs. All these features which may have an appearance similar to a crack have the potential to be misrecognized by a neural network as a crack. Additionally, the complexity of the dataset is increased by corrosion which can be observed in some images. Often, the presence of corrosion is because the paint is intentionally removed around cracks for inspection purposes. In these cases, corrosion makes the crack less visible. Many images were taken using artificial light sources to improve visibility. Examples of images from the CSB dataset and their pixel-wise annotation can be seen in Figure 1. Since, it is computationally inefficient to process entire high-resolution images by a neural network, the images are split into patches of size 512×512 pixels. Furthermore, for our experiments, we pick 1400 patches that contain a crack and 600 patches that do not contain any crack pixels. 10% of all patches were randomly selected for testing of the neural networks, the rest 90% were used for validation and training.

Fig. 1: Examples of images and ground truth crack segmentation from the CSB dataset.