Measurement of hairline cracks using deep learning

Image-based quality assurance

Cracks are to be detected and measured on steel strips with a large data volume of 400 microscope images per sample. The cracks in the image are very thin, approximately 1 pixel wide. In addition, artifacts make crack detection difficult, which is why conventional image processing methods do not offer an exact and robust solution.

Sabrina Fleischanderl et al, “CNN-based crack detection in oxide layers of hot rolled steel sheet samples for the validation of a pickling process model”, Proc. 3rd Symp. on Pattern Recognition and Applications, 2022.

Project requirements

Cracks should not only be detected, but also measured (density, vertical distance in the image)
Large amount of data (400 microscope images per sample)
The cracks are very thin in the image (width ~ 1 px), and other artifacts make crack detection more difficult
Classic image processing cannot provide an exact and robust solution

The solution

Precise and stable segmentation of cracks with neural network (UNET)
60 images were annotated by hand
Wireless connection of each measuring point to the relay station
In addition, a freely available annotated data set (ISBI 2012 Challenge, cell membrane data) was used for training
Precise solution with minimal effort