Multiscale Attention-Enhanced Kolmogorov-Arnold Network for Concrete Porosity Segmentation
Journal of Computing in Civil Engineering · 01.11.2026 · 40 (6)
DOI: 10.1061/jccee5.cpeng-7709 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computing in Civil Engineering, erschienen am 01.11.2026. Verfasst von Alireza Hosseinzadeh und 1 weiteren Autor:innen. Die Arbeit nennt 60 Literaturangaben.
Abstract (Originalsprache)
Accurate porosity quantification in concrete is fundamental to predicting structural durability and service life, yet conventional experimental methods remain destructive, time-intensive, and cost-prohibitive. This paper introduces Atten-UKAN, a novel deep learning architecture that enables fully automated, nondestructive porosity segmentation from computed tomography (CT) imagery. Unlike conventional artificial neural networks that rely on fixed activation functions, the proposed architecture integrates interpretable Kolmogorov–Arnold networks (KANs) with learnable basis spline (B-spline) activation functions, allowing the model to capture complex nonlinear relationships inherent in heterogeneous concrete microstructures. Two critical architectural innovations enable precise handling of substantial pore size variability: (1) a multiscale feature pyramid module that robustly captures features across disparate spatial scales at each encoder stage; and (2) multiscale attention blocks within skip connections that suppress encoder noise while enhancing boundary-relevant features through combined spatial and channel attention mechanisms operating across multiple receptive fields. Validation results demonstrate that Atten-UKAN achieves superior performance with an intersection over union (IoU) of 0.952 and a dice score of 0.975, significantly outperforming state-of-the-art architectures, including U-Net (IoU: 0.923) and TransUNet (IoU: 0.876), under identical conditions. The model exhibits exceptional precision-recall balance (validation precision: 0.982, recall: 0.968), ensuring reliable detection across heterogeneous concrete samples. Fivefold cross-validation confirms robust generalization (mean IoU : 0.943 ± 0.009 ), while ablation studies quantify individual contributions of architectural components. Further, successful validation on historic brick materials demonstrates strong generalization capabilities on unseen material domains. Beyond methodological advances, a user-friendly desktop web application democratizes access to advanced porosity analysis, providing comprehensive pore characterization for civil engineering practitioners.
Zitieren
APA
Hosseinzadeh, A., & Dehestani, M. (2026). Multiscale Attention-Enhanced Kolmogorov-Arnold Network for Concrete Porosity Segmentation. Journal of Computing in Civil Engineering, 40(6). https://doi.org/10.1061/jccee5.cpeng-7709
Vancouver
Hosseinzadeh A, Dehestani M. Multiscale Attention-Enhanced Kolmogorov-Arnold Network for Concrete Porosity Segmentation. Journal of Computing in Civil Engineering. 2026;40(6). doi:10.1061/jccee5.cpeng-7709
BibTeX
@article{Hosseinzadeh2026,
author = {Hosseinzadeh, Alireza and Dehestani, Mehdi},
title = {Multiscale Attention-Enhanced Kolmogorov-Arnold Network for Concrete Porosity Segmentation},
journal = {Journal of Computing in Civil Engineering},
year = {2026},
volume = {40},
number = {6},
doi = {10.1061/jccee5.cpeng-7709}
}
EUVIMED