Material-Condition Fingerprints Enable Scalable, Zero-Label SOM Crack Segmentation via Weight-Vector Transfer

Abstract

Self-Organizing Maps (SOMs) enable label-free and interpretable crack segmentation by clustering pixels in a physically meaningful descriptor space, but their practical use is constrained by a key scalability bottleneck: a SOM must be newly formed for every image. This work removes that bottleneck by introducing a within-material weight-vector transfer mechanism that calibrates once and deploys many times under consistent surface conditions. A single SOM is formed on a representative reference image in the nine-dimensional interpretable descriptor space established in our prior work, and its prototype vectors and internal normalization parameters are stored as a compact material-condition fingerprint. For subsequent images of the same material, segmentation is performed through deterministic nearest-prototype assignment, eliminating per-image SOM formation.

Experiments on aggregate-textured concrete (three classes) and texture-dominated masonry (five classes) show that the transferred SOM closely matches Direct-SOM behavior under strictly zero-label evaluation while yielding substantial runtime reductions that enable full-resolution batch processing. Beyond efficiency, transfer stabilizes decision geometry across repeated inspections and enables rapid screening of crack-free controls without introducing systematic false positives. An auxiliary U-Net refinement trained solely on SOM-generated pseudo-labels further demonstrates that transferred pseudo-labels are sufficiently stable to support downstream spatial regularization without manual annotation.

The proposed mechanism strengthens three deployment-facing advantages—computational efficiency, operational consistency, and zero-label overhead—and highlights limitations under illumination shifts and cross-material use, where re-calibration or condition-specific prototype libraries may be required.