MIR-YOLO: Morphology-Invariant and Reliability-Aware YOLO for Generalizable Building-Surface Defect Detection
International Journal of Advanced Computer Science and Applications · 2026 · 17 (10)
DOI: 10.14569/ijacsa.2026.0171066 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Advanced Computer Science and Applications, 2026. Verfasst von Dongmei Zhao und 1 weiteren Autor:innen.
Abstract (Originalsprache)
Reliable building-surface inspection requires transfer beyond training sites, structures, and acquisition sequences. We evaluate MIR-YOLO, a YOLO11n-derived detector with paired-view appearance diversification and feature/prediction consistency, edge-guided P2–P4 fusion, localization-quality estimation, and source-validation-only isotonic calibration. On the MBDD2025 building-grouped test, mAP50–95 improved from 0.495 to 0.519 over five matched seeds (Δ = 0.024; 4.85% relative; paired Hedges' g ≈ 4.141). The independently confirmed group-hierarchical bootstrap interval for the paired AP gain was [0.015, 0.033]; with 10,000 paired group-label permutations and no null statistic as extreme as observed, Monte Carlo p = 0.000100 (Holm-adjusted p ≤ 0.000400 across four primary questions). Zero-shot CUBIT-Det testing used 1,105 held-out images (718 building, 321 pavement, 66 bridge), increasing pooled external AP from 0.3922 to 0.4250; scenario-level transfer was positive but smaller on the sparsely represented bridge stratum. Like-for-like isotonic calibration reduced grouped LaECE from 0.074 to 0.052. Controlled augmentation-only and per-factor ablations support limited morphology and consistency mechanisms, and additional corruptions include a negative lens-contamination result. At a fixed retrospective 20% image-review budget, error capture improved from 0.635 to 0.685; this does not measure a prospective inspector outcome. Operational effectiveness and structural safety remain outside the current evidence scope.
Zitieren
APA
Zhao, D., & Gu, Y. (2026). MIR-YOLO: Morphology-Invariant and Reliability-Aware YOLO for Generalizable Building-Surface Defect Detection. International Journal of Advanced Computer Science and Applications, 17(10). https://doi.org/10.14569/ijacsa.2026.0171066
Vancouver
Zhao D, Gu Y. MIR-YOLO: Morphology-Invariant and Reliability-Aware YOLO for Generalizable Building-Surface Defect Detection. International Journal of Advanced Computer Science and Applications. 2026;17(10). doi:10.14569/ijacsa.2026.0171066
BibTeX
@article{Zhao2026,
author = {Zhao, Dongmei and Gu, Yu},
title = {MIR-YOLO: Morphology-Invariant and Reliability-Aware YOLO for Generalizable Building-Surface Defect Detection},
journal = {International Journal of Advanced Computer Science and Applications},
year = {2026},
volume = {17},
number = {10},
doi = {10.14569/ijacsa.2026.0171066}
}
EUVIMED