Proceedings of the

The 33rd European Safety and Reliability Conference (ESREL 2023)
3 – 8 September 2023, Southampton, UK

Segmenting Without Annotating: Crack Segmentation and Monitoring via Post-Hoc Classifier Explanations

Florent Foresta, Hugo Portab, Devis Tuiac and Olga Finkd

EPFL, Switzerland.


Monitoring the cracks in walls, roads and other types of infrastructure is essential to ensure the safety of a structure, and plays an important role in structural health monitoring. Automatic visual inspection allows an efficient, cost-effective and safe health monitoring, especially in hard-to-reach locations. To this aim, data-driven approaches based on machine learning have demonstrated their effectiveness, at the expense of annotating large sets of images for supervised training. Once a damage has been detected, one also needs to monitor the evolution of its severity, in order to trigger a timely maintenance operation and avoid any catastrophic consequence. This evaluation requires a precise segmentation of the damage. However, pixel-level annotation of images for segmentation is labor-intensive. On the other hand, labeling images for a classification task is relatively cheap in comparison. To circumvent the cost of annotating images for segmentation, recent works inspired by explainable AI (XAI) have proposed to use the post-hoc explanations of a classifier to obtain a segmentation of the input image. In this work, we study the application of XAI techniques to the detection and monitoring of cracks in masonry wall surfaces. We benchmark different post-hoc explainability methods in terms of segmentation quality and accuracy of the damage severity quantification (for example, the width of a crack), thus enabling timely decision-making.

Keywords: Crack detection, Image classification, Segmentation, Explainable AI, Attribution maps.

Download PDF