Photogrammetry-SfM 3D Reconstruction with 2D Slice-Based YOLOv8 Damage Detection for Architectural Heritage

Abstract

This study proposes an efficient and accurate framework for visualizing, preserving, and restoring architectural heritage by integrating three-dimensional (3D) reconstruction technologies with deep learning-based visual detection algorithms. The objective is to enable intelligent identification and targeted repair of structural defects, thereby advancing the digital conservation of cultural assets. The framework is structured into four layers: data acquisition, 3D modeling, data analysis, and application visualization. In the data acquisition phase, high-overlap image datasets are captured using a GoPro Hero11 action camera. The modeling phase employs the Structure-from-Motion (SfM) algorithm to automatically extract image feature points. Meanwhile, Reality Capture software generates dense point clouds and performs texture mapping—producing high-precision 3D architectural models that retain geometric and textural details. For data analysis, the state-of-the-art You Only Look Once version 8 (YOLOv8) object detection algorithm is applied. The 3D models are sliced and converted into 2D images to detect and locate structural defects such as cracks, spalling, and surface weathering. Experimental results on the validation set demonstrate excellent performance, with an average precision of 96.3%, a recall of 94.7%, and an F1 score of 0.954. The confusion matrix for sectional detection yields diagonal values between 0.81 and 1.00, while classification accuracy for planar structures ranges from 0.91 to 1.00—affirming the model's robustness and real-world applicability. Overall, the proposed method supports high-fidelity reconstruction of architectural structures while enabling precise and automated defect detection via deep learning, providing a reliable quantitative basis for informed and scientific restoration.