Deep learning-based automatic pixel-level detection and severity classification of freeze-thaw damage in concrete dams

Abstract

Freeze-thaw deterioration is a critical durability problem for concrete dams in cold regions, where repeated temperature fluctuation, moisture migration, and surface scaling may progressively weaken concrete cover and threaten long-term structural safety. Although computer-vision methods have been increasingly used for civil infrastructure inspection, most existing studies remain crack-oriented and provide limited support for the simultaneous pixel-level segmentation, physical quantification, and severity grading of multiple freeze-thaw damage morphologies. To address this gap, this study proposes an integrated image-based framework for automatic freeze-thaw damage identification and quantitative severity assessment of concrete dams. A standardized image database is first established through laboratory accelerated freeze-thaw tests on concrete specimens with different strength grades, initial defect states, and cycle stages. An adaptive preprocessing strategy combining pre-denoising, gray-level dynamic clipping, contrast/brightness enhancement, and post-enhancement denoising is then developed to improve the visibility of weak damage features under non-uniform illumination and complex surface textures. On this basis, an improved Mask R-CNN model is constructed by introducing feature-pyramid hierarchy enhancement, multi-scale RoI feature fusion, and a 56 × 56 high-resolution mask branch, enabling instance-level segmentation of spalling, voids, aggregate exposure, and cracks across different spatial scales. The segmented masks are further transformed into engineering-scale damage indicators through camera calibration, including void area ratio, aggregate exposure ratio, spalling area ratio, maximum void diameter, and surface roughness, and k-means clustering is used to establish data-driven severity grades. Experimental validation shows that the proposed preprocessing method increases image contrast by 10.29-27.30 and entropy by 0.24-0.59 while maintaining SSIM values of 0.705-0.876, indicating improved feature enhancement without severe structural distortion. In a field case involving 2,000 dam inspection images, a locally calibrated scale of 0.0913 mm/pixel is used for case-specific physical quantification under comparable imaging geometry, and the extracted spalling area ratios range from 47.06% to 87.95%, consistent with manual engineering inspection records. The k-means severity groups are validated against two independent references: the freeze-thaw cycle count, the standard mechanical deterioration index in the accelerated freeze-thaw testing protocol adopted in this study, and periodic engineering inspection records compiled by qualified dam-safety personnel for the field case, with which the classification results show strong consistency. The main contribution of this work is therefore the engineering-oriented integration of standardized imaging, a purpose-engineered Mask R-CNN architecture, calibrated physical damage indicators, and validated severity interpretation into a single actionable workflow for freeze-thaw damage assessment, maintenance prioritization, and intelligent inspection of concrete dams.