What does this research mean for the field?

Deep learning models for concrete crack detection experience severe performance degradation in real-world field conditions due to domain shift, but this gap can be fully bridged through domain adaptation using very small local datasets on standard CPU hardware. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.SUPPORTS_CONSENSUS.

What question did this study set out to answer?

To evaluate deep learning architectures for concrete crack detection and resolve performance degradation caused by domain shift between benchmark datasets and real-world field inspections.

August 22, 2026Open Access

Automated crack detection in concrete civil infrastructure using deep learning: from benchmark to field deployment with domain adaptation.

View Full Paper

Ask AI

Bookmark

To evaluate deep learning architectures for concrete crack detection and resolve performance degradation caused by domain shift between benchmark datasets and real-world field inspections.
Evaluated three deep learning approaches (a custom CNN, off-the-shelf ResNet50 transfer learning, and partially fine-tuned ResNet50) on the SDNET2018 benchmark using commodity CPU hardware.
Tested benchmark model generalization on 476 real-world inspection images from Île-de-France and performed domain adaptation via fine-tuning on 380 field images.
On SDNET2018, partially fine-tuned ResNet50 achieved 94.1% overall validation accuracy (92.7% on bridge decks), outperforming the custom CNN (90.6%) and standard transfer learning (89.2%).
Direct field deployment of the benchmark model revealed severe domain shift, with accuracy dropping by 79.8 percentage points from 94.1% to 14.3% on 476 field images.
Domain adaptation using 380 field images achieved 100% validation accuracy on 96 field test images after a single training epoch without requiring GPU acceleration.

Implication

Comparative evaluation demonstrates domain adaptation restores deep learning accuracy in field concrete structures, indicating small real-world datasets enable practical deployment.

Abstract

This paper presents a systematic comparison of three deep learning strategies for automated crackdetection in concrete structures, validated against real-world inspection images collected in France.Three models were evaluated on the SDNET2018 dataset: (1) a custom CNN trained from scratch(90.6% validation accuracy); (2) a ResNet50 using off-the-shelf transfer learning (89.2%); and (3)a ResNet50 with layers 3 and 4 fine-tuned, which achieved validation accuracies of 92.7% onbridge decks and 94.1% across the full dataset (D+W+P). Counter-intuitively, pure transfer learningunderperformed the CNN trained from scratch. This is attributed to a domain mismatch: ImageNetderivedfeature representations fail to capture the high-frequency linear texture patterns specific toconcrete cracks. Field validation using 476 images from the Île-de-France region (Essonne, July 2026)revealed a critical domain shift: the v0.4 model’s accuracy plunged from 94.1% on SDNET2018 to just14.3% on field images. This 79.8 percentage-point drop starkly quantifies the gap between laboratorybenchmarks and real-world European operational conditions. Remarkably, domain adaptation viafine-tuning on just 380 field images (model v0.5) fully bridged this gap, achieving 100% validationaccuracy on 96 test images after a single training epoch. This demonstrates that highly compactfield datasets are sufficient for operational deployment. Crucially, all experiments were conductedon standard commodity hardware (CPU) without GPU acceleration. These results establish thetechnical foundation for Metricéa Lab, a platform that integrates AI-driven crack detection intoprofessional structural inspection workflows for bridges, retaining walls, and buildings. The systemaligns directly with the French national framework for structural condition assessment, known asthe IQOA methodology (Image de la Qualité des Ouvrages d’Art — Quality Assessment of CivilEngineering Structures).

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

PAUL BERLIN NDJOKO (2026) studied this question.

synapsesocial.com/papers/6a895f53ca7ade938187dd50 https://doi.org/10.5281/zenodo.22019080

View Full Paper

Ask AI

Bookmark

Implication

Comparative evaluation demonstrates domain adaptation restores deep learning accuracy in field concrete structures, indicating small real-world datasets enable practical deployment.

Automated crack detection in concrete civil infrastructure using deep learning: from benchmark to field deployment with domain adaptation.

Key Points

Implication

Abstract

Discussion

Cite This Study

Automated crack detection in concrete civil infrastructure using deep learning: from benchmark to field deployment with domain adaptation.

Key Points

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading