This research explores how deep learning models can improve the detection of efflorescence in masonry buildings in the Netherlands. Efflorescence, caused by moisture-driven salt transport, poses detection challenges due to its variable appearance, similarity to other surface features, and co-occurrence with other forms of masonry damage.
To address this, the study benchmarked two state-of-the-art object detection models: Mask R-CNN and YOLOv8. On a single-class dataset, Mask R-CNN achieved a peak mAP@0.5 of 0.35, outperforming YOLOv8 (0.30–0.33) in segmentation quality and spatial precision. However, both models suffered from false positives, often misclassifying encrustation, lichens, and graffiti as efflorescence due to visual similarity.
To mitigate this, a multi-class training setup was introduced. Graffiti achieved the highest mAP (0.60) and near-perfect precision due to strong visual contrast, while lichens were classified with high stability. In contrast, efflorescence and encrustation remained difficult to separate, resulting in unstable mAP and precision fluctuations over time. This confirmed that misclassification significantly limits model accuracy when damage types share visual characteristics.
Model performance was further evaluated by incorporating thermal imaging (RGBT), combining aligned RGB and infrared data to detect moisture-driven efflorescence. The RGBT model reduced false positives (as few as 3 per evaluation set) and improved detection confidence, reaching a precision of 0.94 and an average confidence score of 0.96, although it required more epochs to converge and showed increased false negatives in ambiguous scenes. Still, RGBT improved the confidence of the visual detection in real-world, poorly lit, or heritage conditions, where over-segmentation is costly.
A spatial co-occurrence analysis of annotated masks indicated a statistically significant correlation between efflorescence and adjacent damage, supporting the potential of dual-class detection
In conclusion, while deep learning models can support efflorescence detection, especially when enhanced with thermal input and multi-class refinement, their performance depends heavily on dataset quality, annotation strategy, and class separability. These findings offer a foundation for scalable, automated inspection in conservation and diagnostics of masonry.