A method for determining crack depth in reinforced concrete beams through three-dimensional reconstruction

• Damage identification is conducted through three-dimensional reconstruction. • Predict internal damage conditions through surface damage conditions. • A mesoscopic finite element model of full-graded concrete was established. To address the technical challenge of non-destructive internal damage detection in concrete beams—where traditional image-based methods cannot acquire subsurface damage information—this study proposes an intelligent approach that integrates three-dimensional (3D) reconstruction, advanced image analysis, and deep learning techniques. Firstly, to mitigate the impact of image acquisition angles in field inspections on recognition accuracy, a high-precision 3D reconstruction framework was established based on monocular vision and through the triangulation algorithm in the Context Capture software. Compared with the length of the actual object, the average reconstruction error was 1.17%. Combining the YOLOv11 object detection model with U-Net segmentation, as well as skeleton extraction and maximal inscribed circle algorithms, enables precise quantification of crack length and width on beam surfaces, with recognition accuracy rates ranging from 86% to 93.4%. To infer crack depth from surface images, this study investigates the correlation between surface crack width and depth. A comprehensive mesoscopic finite element model, incorporating aggregates, mortar, and interfacial transition zones (ITZs), was developed and experimentally validated (modeling error below 15%). Through parameterized fracture analysis, an extensive database containing over 800 crack propagation scenarios was generated. Based on this dataset, a deep residual neural network prediction model—optimized via a cyclic learning rate strategy—was established to predict crack depth from surface characteristics, yielding an average error of approximately 4% relative to beam depth. Field validation demonstrated strong agreement between the predicted results and those obtained by ultrasonic testing. The proposed method offers a reliable, efficient, and non-destructive solution for estimating internal crack depth solely from surface images, exhibiting substantial promise for large-scale application in engineering diagnostics, especially in safety-critical structural assessments.