Integrating Close-range Photogrammetry With Deep Learning for Intelligent 3D Documentation of Cracks in Wind-rain Bridge Heritage

Abstract

To foster the in-depth application of real-scene 3D technology in architectural heritage conservation, this study focuses on the wind-rain bridges along the Xiao-He Ancient Road (China) and conducts research on a visualization method for apparent cracks based on real-scene 3D modeling. High-resolution images of the bridges were obtained through close-range photogrammetry, enabling the construction of a detailed real-scene 3D model. Using established algorithms such as DeepLabV3+ and YOLO, cracks in the architectural images of the wind-rain bridges were identified. Leveraging the multi-view overlap of the close-range images, the crack detection results from the deep learning algorithms were accurately located within the real-scene 3D model of the bridges. On this basis, the real-scene 3D model and the crack identification results were integrated into the Unity3D platform, achieving visualization of apparent cracks in the wind-rain bridge. The proposed method is directly applicable to heritage surveys, providing technical support for compiling digital archives of apparent cracks. These archives are of significant value as they can directly inform future conservation planning and monitoring strategies.