From Classification to Segmentation with Explainable AI: A Study on Crack Detection and Growth Monitoring

XAI can reduce labeling costs for infrastructure monitoring: The paper demonstrates that explainable AI methods can generate useful segmentation masks from image-level labels alone, rather than requiring expensive pixel-by-pixel annotations, making automated crack detection more practical for real-world infrastructure inspection.

XAI-derived segmentations are "good enough" for practical purposes: While the segmentation quality from XAI methods is lower than fully supervised approaches, the paper finds these masks remain "meaningful and enable severity monitoring," suggesting a practical trade-off between annotation effort and acceptable performance.

The approach enables crack growth tracking over time: By using XAI to monitor crack severity and progression, the methodology supports long-term structural health monitoring without the need for continuous expensive manual annotation of new images.

Different XAI methods produce varying results: The paper evaluates "various XAI methods," indicating that the choice of explainability technique matters for segmentation quality—suggesting practitioners need to carefully select which XAI approach fits their specific infrastructure monitoring needs.

Automation addresses safety and accessibility challenges: The paper emphasizes that automatic visual inspection is "especially effective in hard-to-reach areas," meaning this AI approach can reduce human safety risks by automating inspections of dangerous or difficult-to-access infrastructure locations.