Semantic Segmentation for Hot Cracking

The images in this dataset are light microscope images of the surfaces of modified Varestraint-Transvarestraint test samples. The training images show the surface of a tungsten inert gas weld with possible hot cracks. The dataset is intended for use in training crack segmentation networks. This database consists of images in .png format. The training images were taken in 24-bit, 3-color channels with an optical light microscope (Keyence). The corresponding mask images are 8-bit grayscale images, where 255-valued pixels (white) denote crack areas and 0-valued pixels (black) denote non-cracked areas. This dataset was created manually with the help of a self-written segmentation program based on OpenCV. I used it to train networks based on ViT, ConvNeXt, U-Net, and EfficientNet architectures. The images were taken at magnifications of 50x and 100x. The dataset is structured as follows: Wire → Magnification (50, 100) → Data (train, mask = label) → Numbered images (1, 2, 3, …) (There may be missing numbers. The reasons are: 1) I did not take images of all surfaces if they were not cracked, and 2) the 100x magnification images were taken at a later point, by which time some samples had already been prepared for further evaluation, so no images could be taken.) For every image number in train, there is an equivalent number in mask with the same image size (but a different channel size, since train is 3-channel color and mask is single-channel). There is no fixed image size (h × w). If you want to train a network, I recommend either rescaling or taking sub-images for both the training and mask sets, or using architectures that can handle varying input sizes. The research project " Evaluation of design-induced solidification cracks during welding of high-strength fine-grained structural steels" is funded by the Federal Ministry of Economics and Climate Protection as part of the "Industrial Collective Research" programme on the basis of a resolution of the German Bundestag. This project IGF 01IF21837N FE 1 / P 1516 from the Research Association for steel Application (FOSTA), Düsseldorf, is carried out at Bundesanstalt für Materialforschung und -prüfung.