Answers are grounded in this paper and cite their sources.
Systematic computer vision method monitors crack evolution in heritage buildings, supporting preservation efforts.
To meet the high-precision and non-contact requirements of crack monitoring in heritage buildings, this study proposes a systematic approach that integrates deep learning, image stitching, and skeleton-based geometric analysis for crack detection and temporal evolution monitoring. Jiangzhuang Garden, a heritage site by West Lake in Hangzhou, China, is used as the case study, with multi-temporal crack images acquired and converted to millimeter-scale measurements via camera calibration and scale computation. Three deep learning models, U 2 -Net, Mask R-CNN, and DeepCrack, are compared in terms of training performance and prediction quality. DeepCrack, demonstrating superior connectivity and stability, is selected as the primary crack extraction method. Post-processing includes connected-component analysis and skeletonization for noise removal, followed by pointwise width measurement using quadratic fitting and normal-direction search. For long cracks, a SIFT- and RANSAC-based stitching pipeline is implemented to generate complete crack panoramas. Temporal monitoring is achieved by aligning multi-period images using a sliding-window matching algorithm, and width variation is analyzed through mean, standard deviation, and threshold-based evaluation within a unified coordinate system. Results show that the proposed framework achieves robust crack detection and stable temporal comparison in complex backgrounds, providing technical support for long-term monitoring, digital management, and damage assessment of heritage buildings. • Propose vision-based framework for crack monitoring in heritage buildings. • Integrate deep learning, image stitching, and skeleton-based analysis. • Achieve millimeter-scale crack measurement via camera calibration and scaling. • Enable multi-temporal crack evolution monitoring with sliding-window matching.
No takes yet. Share an insight, caveat, or question.
Ye et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: