From Semi-Automatic to Fully Automatic: An Auditable SOM Pipeline for Label-Free Crack Segmentation

Abstract

Self-Organizing Maps (SOMs) enable unsupervised, label-free, and interpretable crack segmentation, yet practical deployment is often limited by manual choices of feature-space dimension, cluster number, and visual crack-class confirmation. This paper presents a fully automatic and auditable SOM-based pipeline that removes these expert-dependent degrees of freedom while preserving interpretable pixel descriptors. The method forms a deterministic decision chain integrating PCA-based automatic dimension selection, elbow-based cluster-number determination, an intensity-driven crack-class selector with signature-based auditing evidence (radar and bar summaries), and fixed post-processing to produce a binary crack mask without any pixel-wise labels. An optional per-image CNN refinement stage, trained only on SOM pseudo-labels, serves as a lightweight spatial regularizer to refine boundary coherence. Experiments on three representative field conditions—aggregate-textured concrete, texture-dominated masonry, and shadow-affected surfaces—demonstrate robust end-to-end automation, high SOM–CNN internal consistency (Dice 0.86–0.91), and practical CPU-only runtimes (4–14 min per image). The results position fully automatic, interpretable clustering as a viable crack segmentation pathway for edge-oriented inspection workflows under realistic constraints and distribution shifts.