Structural Crack Detection

Automated Multi-Class Structural Damage Detection and Segmentation for UAV-Based Infrastructure Inspection Using Deep Learning

Reference implementation for the paper submitted to Structures (Elsevier, 2026).

A three-stage deep learning pipeline for automated structural damage detection, segmentation, and UAV inspection simulation. Targets bridges, offshore platforms, and other civil infrastructure.

---

Overview

Stage	Task	Model	Key Result
1	Four-class damage classification (crack, corrosion, spalling, intact)	ResNet50 with differential learning rates	85.69% accuracy
2	Pixel-level semantic segmentation	DeepLabV3+ with ResNet50 backbone	mIoU 0.789 (3-class), 0.645 (4-class)
3	UAV grid inspection simulation	Boustrophedon traversal over 4×5 grid	16/20 zones detected (80%)

All stages were trained and benchmarked on CPU hardware (Intel processor, no GPU). Measured Stage 1 inference rate: 61 FPS.

---

Repository Structure

structural-crack-detection/
├── src/                      # Core model and pipeline code
├── results/
│   └── plots/                # Publication-ready figures
├── analyze_crack_masks.py    # 4-class segmentation analysis
├── analyze_labels.py         # Stage 2 segmentation evaluation
├── augment_corrosion.py      # Targeted augmentation for corrosion class
├── augment_spalling.py       # Targeted augmentation for spalling class
├── drone_simulation.py       # Stage 3 UAV grid inspection simulation
├── extract_classes.py        # Class extraction from RGB masks
├── failure_analysis.py       # Qualitative failure case analysis
├── fix_gradcam.py            # Grad-CAM visualization
├── gradcam.py                # Grad-CAM core implementation
├── inference_time.py         # CPU inference benchmarking (Stage 1)
├── per_class_iou.py          # Per-class IoU reporting
└── README.md

---

Requirements

• Python 3.9+
• PyTorch 2.1.0
• torchvision
• numpy, matplotlib, scikit-learn, Pillow, opencv-python

Install with:

pip install -r requirements.txt

---

Datasets

Three publicly available datasets are used across the three stages:

1. SDNET2018 — 56,092 concrete surface images (bridge decks, walls, pavements). Binary crack annotations. Used for Stage 1 (crack and intact classes). Available from the original dataset authors.

2. Corrosion and Spalling Segmentation Dataset (Raidathmane, Kaggle) — 1,580 image–mask pairs with pixel-level RGB annotations. Used for Stage 1 (corrosion and spalling classes) and Stage 2 (3-class segmentation). https://www.kaggle.com/datasets/raidathmane/corrosion-and-spalling-concrete-defect-segmentation

3. UAV-Based Crack Detection Dataset (Ziya07, Kaggle) — 315 real UAV images of cracked concrete surfaces with binary segmentation masks. Used for Stage 2 (4-class segmentation). https://www.kaggle.com/datasets/ziya07/uav-based-crack-detection-dataset

Datasets are not bundled with this repository. Follow the source links to download.

---

Pretrained Model Weights

Pretrained weights for all three models are available as assets in the v1.0 Release:

https://github.com/far-reach/structural-crack-detection/releases/tag/v1.0

File	Stage	Description	Size
best_resnet50_multiclass.pth	Stage 1	ResNet50 four-class classifier	92 MB
best_deeplabv3_segmentation.pth	Stage 2	DeepLabV3+ three-class segmentation	161 MB
best_deeplabv3_4class.pth	Stage 2	DeepLabV3+ four-class segmentation (includes UAV crack data)	161 MB

Place downloaded weights in a local models/ directory (create if absent) before running inference scripts.

---

Results

Stage 1 — Four-class classification

• Overall accuracy: 85.69% on original held-out validation set (1,600 samples)
• Per-class F1: crack 0.839, intact 0.862, corrosion 0.860, spalling 0.886
• Differential learning rate strategy improved accuracy by 8.25% over frozen-backbone baseline

Stage 2 — Semantic segmentation

• 3-class (background / corrosion / spalling): mIoU 0.789, pixel accuracy 0.918
• 4-class (+ crack): mIoU 0.645, crack IoU 0.206 (reflecting 1.78% mean crack pixel density)

Grad-CAM Interpretability

Grad-CAM visualizations confirm the classifier attends to visually relevant damage regions across all four classes.

Failure Analysis

Three representative misclassification cases are included to identify targeted improvement directions (hairline cracks on textured surfaces, wet corrosion, and boundary spalling).

Stage 3 — UAV grid inspection simulation

• 16 of 20 damaged zones detected (80%)
• Mean inference confidence: 88.4%
• Total CPU inference time for 20 frames: 0.33 s

Computational performance (Stage 1, CPU)

• Mean inference time: 16.29 ms per frame
• CPU throughput: 61.39 FPS
• Exceeds typical UAV camera capture rate (25–30 FPS)

---

Citation

If you use this code or the pretrained weights in your work, please cite the paper:

Abtahi, S.F. (2026). Automated Multi-Class Structural Damage Detection and
Segmentation for UAV-Based Infrastructure Inspection Using Deep Learning.
Structures. [Manuscript under review.]

Citation will be updated upon acceptance.

---

License

MIT License — see LICENSE for details.

---

Author

Seyed Farhad Abtahi Independent Researcher, Istanbul, Türkiye ORCID: 0009-0007-8865-2300 Contact: farhadabtahi91@gmail.com