Deep Learning-Based 3D Reconstruction of Interior Surface Defects in Post-Earthquake Buildings

Frequent moderate to minor earthquakes often cause significant structural damage to buildings, manifested as cracks and spalling on interior surfaces. These defects compromise structural integrity and can lead to catastrophic collapses during aftershocks, posing significant risks to rescue personnel. Therefore, promptly assessing interior defects is vital for preventing secondary injuries. This study explores the utilization of selfdeveloped ejectable vision sensors to capture images of interior surface defects, combined with deep learning algorithms to automatically identify defects and map their spatial distribution. However, existing 3D reconstruction methods struggle to produce accurate depth maps in scenes with in-place rotation and posefree conditions. To address this challenge, this study proposes a novel 3D reconstruction model for interior surface defects in postearthquake buildings. Leveraging monocular depth estimation as a foundational prior, the proposed model is equipped with a camera parameter normalization module to produce accurate depth maps. Additionally, a camera pose joint optimization module is developed to ensure the global alignment of depth maps and pose estimation across varied camera parameters. Furthermore, the integration of a point cloud fusion module facilitates the detailed reconstruction of post-earthquake building interior scenes. The effectiveness of this proposed model has been confirmed through extensive experimental evaluations in laboratory environments and simulated defect scenes.