CrackMNIST - A Large-Scale Dataset for Crack Tip Detection in Digital Image Correlation Data

Introduction

Fatigue crack growth (FCG) experiments play a crucial role in materials science and engineering, particularly for the safe design of structures and components. However, conventional FCG experiments are both time-consuming and costly, relying primarily on integral measurement techniques such as the potential drop method to determine crack length.

Digital Image Correlation (DIC) is a non-contact optical technique that enables full-field displacement measurements during experiments. Accurately identifying crack tip positions from DIC data is essential but challenging due to inherent noise and artifacts.

Recently, a deep learning-based approach was introduced to automatically detect crack tip positions [1,2]. This method involved manually annotating a single experiment to train a convolutional neural network (CNN). Furthermore, an iterative crack tip correction technique was later developed to enhance detection accuracy [3]. However, this method is not fully automated and requires more time than applying a pre-trained CNN. With the rise of self-driven laboratories generating vast amounts of DIC data [4,5], reliable crack tip detection is essential for efficient and rapid data evaluation.

References:

1. Strohmann T et al. (2021) Automatic detection of fatigue crack paths using digital image correlation and convolutional neural networks. Fatigue and Fracture of Engineering Materials and Structures 44: 1336-1348 https://doi.org/10.1111/ffe.13433
2. Melching D et al. (2022) Explainable machine learning for precise faticue crack tip detection. Scientific Reports 12, 9513 https://doi.org/10.1038/s41598-022-13275-1
3. Melching D et al. (2024) An iterative crack tip correction algorithm discovered by physical deep symbolic regression. International Journal of Fatigue, 187, 108432 https://doi.org/10.1016/j.ijfatigue.2024.108432
4. Paysan F et al. (2023) A Robot-Assisted Microscopy System for Digital Image Correlation in Fatigue Crack Growth Testing. Experimental Mechanics, 63, 975-986 https://doi.org/10.1007/s11340-023-00964-9
5. Strohmann T et al. (2024) Next generation fatigue crack growth experiments of aerospace materials. Scientific Reports 14, 14075 https://doi.org/10.1038/s41598-024-63915-x

Objective

The objective of this project is to create a diverse, large-scale, and standardized dataset designed for the training and evaluation of deep learning-based crack tip detection methods. In addition to supporting research and practical applications, the dataset aims to serve an educational purpose by providing a high-quality resource for students and researchers in the field of material science and mechanics.

DIC data

The dataset contains DIC data in the form of planar displacement fields ($u_x, u_y$) both measured in $mm$ from eight FCG experiments performed on different materials and specimen geometries. The tested materials (AA2024, AA7475 and AA7010) are aluminum alloys with an average Young's modulus (E) of approximately 70 GPa and a Poisson’s ratio (ν) of 0.33. For details, please refer to the corresponding data sheets.

The applied maximum nominal uniform stress for MT-Specimen is σ_N is 47 MPa (sinusoidal loading, constant amplitude). The minimum load can be derived from R=F_min/F_max. The expected Stress Intensity Factors K_I vary approximately between 1 and 40 MPa√m.

Experiment	Material	Specimen Type	Thickness [mm]	Orientation	R
MT160_2024_LT_1	AA2024r	MT160	2	LT	0.1
MT160_2024_LT_2	AA2024r	MT160	2	LT	0.3
MT160_2024_LT_3	AA2024r	MT160	2	LT	0.5
MT160_2024_TL_1	AA2024r	MT160	2	TL	0.1
MT160_2024_TL_2	AA2024r	MT160	2	TL	0.3
MT160_7475_LT_1	AA7475r	MT160	4	LT	0.1
MT160_7475_TL_1	AA7475r	MT160	4	TL	0.3
CT75_7010_SL45_1	AA7010f	CT75	12	SL45°	0.1

---

^r Rolled Material ^f Forged Material

Data annotation

Crack tip positions in the DIC data are annotated with the high-fidelity crack tip correction method from [3] (see Figure below).

The crack tip positions are stored as binary segmentation masks such that the labelled datasets can directly be used for training semantic segmentation models.

Labelled datasets

We provide three datasets of different sizes ("S", "M", "L"). The datasets are split into training, validation, and test sets. The following table shows the number of samples in each dataset.

Dataset	Training	Validation	Test
S	10048	5944	5944
M	21672	11736	11672
L	42088	11736	16560

The datasets are provided in four different pixel resolutions ($28 \times 28$, $64 \times 64$, $128 \times 128$, $256 \times 256$) and stored in HDF5 format.

An overview which experiment is included in which dataset for training, validation and testing can be found in the file size_splits.json.

Visualization of labelled data samples

The following figure shows examples of labelled data samples from the CrackMNIST dataset.

The inputs consist of the planar displacement fields ($u_x, u_y$), and the outputs are the binary segmentation masks.

Visualization of different pixel resolutions

The figure below shows the y-displacement field of a DIC sample at different pixel resolutions.

Usage

Installation

The package can be installed via pip:

pip install crackmnist

Datasets are uploaded to Zenodo and are downloaded automatically upon usage.

Getting started

The datasets can be loaded using the implemented class CrackMNIST as follows

from crackmnist import CrackMNIST

train_dataset = CrackMNIST(split="train", pixels=28, size="S")

Here, the parameters split, pixels, and size specify the dataset split, and pixel resolution, respectively.

The folder examples contains a Jupyter notebook getting_started.ipyb that demonstrates how to train a simple U-Net segmentation model, and evaluate and visualize the results.

Contributors

Code implementation and data annotation by:

• Erik Schultheis
• David Melching

Experiment conduction and DIC data acquisition by:

• Florian Paysan
• Ferdinand Dömling
• Eric Dietrich

Supervision and conceptualization by:

• David Melching
• Eric Breitbarth

Citation

If you use the dataset or code in your research, please cite this GitHub repository:

@misc{crackmnist,
  title={Crack-MNIST - A Large-Scale Dataset for Crack Tip Detection in Digital Image Correlation Data},
  author={David Melching and Erik Schultheis and Florian Paysan and Ferdinand Dömling and Eric Dietrich and Eric Breitbarth},
  journal={GitHub repository},
  howpublished={\url{https://www.github.com/dlr-wf/crackmnist}},
  year={2025}
}

License and Limitations

The package is developed for research and educational purposes only and must not be used for any production or specification purposes. We do not guarantee in any form for its flawless implementation and execution. However, if you run into errors in the code or find any bugs, feel free to contact us.

The code is licensed under MIT License (see LICENSE file). The datasets are licensed under Creative Commons Attribution 4.0 International License (CC BY 4.0).