Unsupervised Semantic Segmentation for Crack Detection in Concrete Elements

Reliable and early detection of cracks is crucial for assessing the condition and maintaining structures. Thus, numerous methods were employed to detect cracks. Semantic segmentation methods have demonstrated superior results in detecting cracks in comparison to conventional methods (e.g. traditional image-based techniques). However, training such networks requires human-annotated datasets. To eliminate the need for labor-intensive pixel-annotation of the images, unsupervised semantic segmentation offers an alternative solution. In this research, two state-of-the-art models, STEGO and EAGLE, due to their success on benchmarks, were chosen to be trained on the microCT and DeepCrack datasets. The quantitative results of STEGO trained on microCT are promising, but the qualitative results show that it failed to detect cracks, whereas EAGLE with nearly identical metrics could partially detect cracks. Based on the qualitative results on DeepCrack, both models detected cracks in this dataset, however, EAGLE performed worse than STEGO in metrics.