Unsupervised and noise-resilient crack detection in concrete and masonry via self-organizing maps

This paper proposes a lightweight, interpretable, and label-free crack segmentation framework based on unsupervised Self-Organizing Map (SOM) clustering with minimal human-in-the-loop semantic labeling at the cluster level. Unlike prior SOM-based approaches that rely on patch-level descriptors or handcrafted clustering heuristics, our method performs direct pixel-level segmentation by clustering low-level features—including grayscale, contrast, edge sharpness, thinness, hue, local averages, and relative grayscale indicators—without requiring any labeled data, model training, or manual thresholds. Tested on field-acquired imagery from diverse surface types—including concrete, brick, and shadow-affected regions—the framework achieves 2 % error on textured brick and 4–5 % on concrete and shadowed surfaces, consistently outperforming classical filters such as Canny (up to 15 %) and Sobel (up to 12 %). Feature ablation experiments further reveal material-specific sensitivities, offering insights into SOM’s internal decision process and reinforcing its interpretability and robustness. While Sobel performs comparably on smooth concrete with clear edges, its accuracy deteriorates substantially under textured or low-contrast conditions, whereas SOM maintains stable performance across all materials under the same evaluation protocol. To assess downstream utility, SOM-generated masks were used as optional pseudo-labels to supervise a compact U-Net trained on only three images, which reproduced accurate segmentations within the same domain. This demonstrates SOM’s capacity to serve as a cost-free pseudo-labeling mechanism for weakly supervised SHM workflows in low-data regimes. Designed for edge deployment, the system processes 276 × 434–800 × 600 grayscale images in 8–12 s on standard CPUs, requiring no GPU or retraining—making it well-suited for UAV-based bridge inspections and other field scenarios where compute resources and data are limited. This hybrid paradigm—merging unsupervised clustering with lightweight deep learning—offers a largely training-free, label-free, and explainable alternative to data-hungry models for infrastructure diagnostics under real-world constraints.