Noise Resilient Concrete and Masonry Crack Detection Using Self-Organizing Maps

This paper presents a lightweight, training-free framework for automated crack detection in concrete and masonry 7 structures using Self-Organizing Maps (SOMs). By clustering pixel-level features—including grayscale, edge 8 gradients, contrast, hue, and thin-crack indicators—the method performs unsupervised segmentation without labeled 9 data, manual thresholds, or deep models. It demonstrates strong robustness to noise, shadows, and textured surfaces 10 across field-acquired imagery. Feature ablation experiments reveal material-dependent contributions of each input, 11 offering insights into SOM’s internal structure and interpretability. In addition, SOM-derived pseudo-labels are used 12 to train CNNs, achieving high segmentation accuracy and generalization across unseen domains. Designed for UAV 13 and mobile SHM platforms, the framework operates with low memory (<100 MB), fast inference (<30 s/image), and 14 minimal human input—offering an interpretable, scalable alternative to heuristic filters and data-intensive networks 15 for real-world infrastructure diagnostics.