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

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

We provide three datasets of different sizes ("S", "M", "L"). The datasets are split into training, validation, and test sets.The digital image correlation (DIC) data consists of full-field planar displacements in x- and y- direction and is interpolatedto a regualar image-like grid. The data is provided in different resolutions: 28x28, 64x64, 128x128 and 256x256Due to data storage constraints, for the highest resolution 256x256 the sizes "M" and "L" are not provided here butare available upon request.The CrackMNIST datasets can be easily loaded using the Python package crackmnist.For further information, we refer to https://github.com/dlr-wf/crackmnist