Structural health monitoring (SHM) remains underutilized in dam engineering practice despite its potential to enhance asset management through earlier damage detection and risk-informed decision-making. To advance research in automated crack and spalling detection, we introduce DamSegment , a high-resolution image dataset curated to support development of deep learning-based damage assessment of large-scale concrete infrastructure. The dataset comprises 3,500 image patches derived from images captured during an on-site visit of a concrete dam. Unlike existing public datasets, DamSegment includes dam-specific deterioration mechanisms, such as freeze-thaw-induced spalling and complex map-cracking patterns, which are often underrepresented in general concrete damage datasets. Images were acquired using multiple mobile devices to ensure diverse viewpoints and surface conditions. The collected images were divided into 512 × 512-pixel patches and presented in a manner that supports damage classification, detection, and segmentation tasks. Images intended for damage segmentation are sorted by crack detection difficulty, enabling curriculum learning-based model training in which neural networks progressively learn from simple to complex damage patterns. Each image is annotated with pixel-level polygons and masks, making the dataset suitable for multiple computer vision tasks including damage classification, detection, and segmentation.
Gharehbaghi et al. (Sun,) studied this question.