Annotated Image Dataset of Masonry Building Cracks for Training Machine Learning Models for Crack Segmentation

This dataset is intended for training a Python-based machine learning model for masonry crack segmentation. It comprises 100 original images of masonry cracks collected during structural inspections of low‑rise residential properties in the UK, with corresponding pixel-level crack masks. All images were taken using standard mobile phone or compact camera devices under typical site conditions. The dataset is representative of real-life inspection environments, incorporating various masonry types, crack presentations, lighting conditions and background details. House numbers, street names and other potential sources of personally identifiable information have been manually pixelated for compliance with UK GDPR. EXIF metadata has been removed from the image files. Each image is paired with a binary segmentation mask that identifies the pixel-level crack region. The masks were generated through manual annotation using a digital image-editing tool. Crack pixels are labelled as the positive class (white, value = 255), with all other regions assigned as background (black, value = 0). Annotation was undertaken manually to ensure high accuracy delineation of crack boundaries. These masks form the ground truth for training of CNN-based segmentation models. The dataset uses a one‑to‑one naming convention to link each image file with the corresponding mask file. The filenames match, with the mask file appending the suffix _mask. For example, image: masonry_crack_001.jpg is paired with masonry_crack_001_mask.png. This pattern is consistent for all 100 image–mask pairs.