Mask-guided stereo vision technique for crack depth estimation of concrete bridge deck

Abstract

Durability of reinforced concrete structures depends on early detection and accurate assessment of damage that accelerates reinforcement corrosion and long term deterioration. Cracks on bridge decks allow moisture and chloride ingress, making crack depth an important indicator of deterioration severity that cannot be captured by conventional visual inspection or two dimensional image based methods. Although stereo vision provides a cost effective approach for three dimensional surface reconstruction, reliable depth estimation in narrow cracks remains challenging due to weak texture, low contrast, and smoothness assumptions in disparity matching. To address this limitation, this study proposes a mask guided stereo vision framework that integrates semantic crack localization with geometry based depth reconstruction. A transfer learning based U-Net model generates crack masks from high resolution stereo images, which guide a crack aware Semi Global Block Matching (SGBM) process for disparity estimation. Experiments on controlled specimens and real bridge imagery show improved disparity coherence and more reliable crack depth estimation compared with conventional SGBM. The proposed framework enables scalable non contact crack depth assessment for structural health monitoring (SHM) and maintenance of concrete infrastructure.