PAGG-Net: Physically Aligned High-Resolution Concrete Crack Dataset

This dataset contains high-resolution concrete surface images with pixel-wise crackannotations and four derived multi-scale tile databases used to train and evaluate PAGG-Net (Physically Aligned Glance–Gaze Network). Each original image (4224×5632 pixels) ispaired with a binary crack mask, and is tiled into strictly aligned multi-scale patchesat physical scales 1×, 2×, 4×, 8×, and 16×. The databases include detailed CSV mappingsand MD5 hashes for reproducibility, and preserve an image-level train/val/test split.The data support research on crack segmentation, multi-scale fusion, tiled inference,and automated infrastructure inspection.