A dataset for semantic segmentation of early-stage cracks in railway sleepers

The publicly available dataset, developed to facilitate research on the detection and semantic segmentation of early-stage cracks in concrete railway sleepers, is distributed in two complementary representations: dataset_source, containing the original full-resolution images, and dataset_224, comprising 224 × 224-pixel patches derived from these images. Both representations originate from the same collection of 209 railway sleeper images acquired under real-world operating conditions on an active railway line in Poland. The source dataset includes 33 images showing visible cracks and 176 images without annotated cracks, while the tiled representation consists of 31,757 image–mask pairs, including 891 positive and 30,866 negative patches. These two representations are intended to provide flexibility in data handling and model development and should not be considered independent datasets.The original images depict entire railway sleepers rather than isolated damage regions, thereby preserving the spatial context of defects within the concrete structure. Their dimensions range from 344 to 712 pixels in width and from 3,344 to 5,001 pixels in height. The images were decoded from JPEG files and saved as lossless RGB PNG files without resizing or modification of the decoded pixel values.Both representations include corresponding single-channel, pixel-level segmentation masks. The masks use a binary encoding, with pixel values of 0 representing the background and 255 indicating the presence of a crack. Images containing cracks have at least one annotated foreground pixel, whereas negative samples are accompanied by entirely zero-valued masks. The annotations were verified by a railway civil engineer with expertise in sleeper damage assessment. The dataset focuses on early-stage cracks, including defects that may be difficult to identify through visual inspection but can provide relevant diagnostic information for assessing the technical condition of railway infrastructure.In dataset_224, the original images and their corresponding masks are divided into 224 × 224-pixel patches to facilitate their use in deep-learning workflows and standard image-processing pipelines. The tiles were generated using a nominal stride of 112 pixels in both horizontal and vertical directions, resulting in a 50% overlap between adjacent patches under regular sampling. Additional adjustments to the final tile positions ensure complete coverage of the original image boundaries. No resizing or padding was required for the source images.Both dataset representations include machine-readable metadata describing image identifiers, class labels, dimensions, crack pixel counts and relationships between individual patches and their corresponding source images. The tiled dataset additionally provides information on patch coordinates and valid image regions, while the accompanying reconstruction scripts enable the original full-resolution images and segmentation masks to be recovered. Reconstruction is performed using the recorded patch positions and image dimensions, with overlapping regions checked for pixel consistency. The reconstructed images and masks reproduce the original decoded RGB pixel values and annotation masks exactly. Thus, the tiled representation provides flexibility for model development without losing access to the spatial context and information contained in the original images.Neither representation includes a predefined training, validation or test split. When preparing subsets for model development and evaluation, all patches belonging to the same group should be assigned to a single subset to reduce the risk of data leakage resulting from overlapping image regions. The original-image and tiled representations constitute two forms of the same dataset and should not be treated as independent collections for training and evaluation. Furthermore, the dataset does not provide spatial calibration, meaning that pixel measurements cannot be directly interpreted as physical crack dimensions.By providing full-sleeper imagery, pixel-level annotations and a readily usable tiled representation with the possibility of exact image reconstruction, this dataset addresses the limited availability of publicly accessible resources for early-stage railway sleeper crack segmentation. It is intended to support the development, training and evaluation of automated computer-vision methods for railway infrastructure inspection, condition assessment and diagnostics.