Back to datasets
Dataset assetOpen Source CommunityStructural Health MonitoringSteel Structure Damage Detection
Steelcrack Dataset
The Steelcrack dataset consists of images captured directly from steel structures across various projects. It includes 3,300 training images, 525 validation images, and 530 test images, each sized 512 × 512 pixels. Some images originate from the first International Structural Health Monitoring Competition, while the remainder were provided by the authors and re‑annotated for finer granularity.
Source
github
Created
Mar 22, 2023
Updated
Apr 1, 2024
Signals
513 views
Availability
Linked source ready
Overview
Dataset description and usage context
Dataset Overview
Dataset Name
- Steelcrack Dataset
Dataset Download
- Download links: Google Drive or OneDrive
Basic Information
- Image size: 512 × 512
- Training set: 3,300 images
- Validation set: 525 images
- Test set: 530 images
- Image source: Partially from the 1st International Structural Health Monitoring Competition, the rest provided by the authors. All images have been re‑annotated for finer detail.
Experimental Results
| Method | mi IoU (%) | mi Dice (%) | #Params (M) | MACs (G) |
|---|---|---|---|---|
| U‑Net | 68.49 | 75.13 | 7.77 | 55.01 |
| U‑Net++ | 72.23 | 78.37 | 9.16 | 138.63 |
| Attention U‑Net | 71.25 | 77.54 | 34.88 | 266.54 |
| CE‑Net | 76.00 | 81.54 | 29.00 | 35.60 |
| DeepLabv3+ (MobileNetv2) | 68.22 | 71.07 | 5.81 | 29.13 |
| DeepLabv3+ (Xception) | 67.40 | 71.48 | 54.70 | 83.14 |
| DeepLabv3+ (ResNet‑101) | 69.04 | 69.45 | 59.34 | 88.84 |
| SCRN | 73.23 | 78.91 | 25.23 | 31.92 |
| TransUNet | 64.34 | 72.55 | 67.87 | 129.96 |
| CrackSeU‑B | 70.42 | 80.50 | 3.19 | 11.22 |
| CrackSeU‑L | 71.66 | 81.24 | 4.62 | 28.22 |
| DconnNet | 74.73 | 83.40 | 28.38 | 24.79 |
| BGCrack V1 | 77.16 | 85.33 | 2.32 | 15.76 |
Citation Information
- BibTeX citation:
@article{HE2024BGCrack,
title = {Crack segmentation on steel structures using boundary guidance model},
journal = {Automation in Construction},
volume = {162},
pages = {105354},
year = {2024},
issn = {0926-5805},
doi = {https://doi.org/10.1016/j.autcon.2024.105354},
url = {https://www.sciencedirect.com/science/article/pii/S0926580524000906},
author = {Zhili He and Wang Chen and Jian Zhang and Yu-Hsing Wang},
keywords = {Crack inspection, Deep learning, Boundary guidance method, Benchmark dataset}
}
Need downstream help?
Pair the dataset with AI analysis and content workflows.
Once the source passes your review, move straight into summarization, transformation, report drafting, or presentation generation with the JuheAI toolchain.