Three-Dimensional (3D) Semantic Reconstruction of Complex Historical Buildings Based on Multi-Source UAV Images and Deep Learning GAN Model

Existing reconstruction methods commonly face practical bottlenecks such as ambiguous feature matching, disjointed geometric semantic information, and insufficient controllability of generated texture styles when dealing with programmatic complex historical buildings. Therefore, it is necessary to conduct research on component-level 3D semantic reconstruction technology to address these pain points. To improve the 3D semantic reconstruction of complex historical buildings, this study proposes a method based on multi-source UAV images using deep learning via a generative adversarial network (GAN). The Application Domain Extension mechanism of CityGML is adopted to construct the semantic model. Components including brackets, beams, and columns, along with their modular-scale relationships and spatial topology, are defined. A UAV collects multi-angle, multi-route images. The scale of feature point detection and the range of depth prediction are constrained using component parameters. Geometric features of point clouds, such as linearity, flatness, and verticality, are extracted, and feature fusion is conducted via the PointNet network. Superpoint graphs are constructed to retain topological relationships. A conditional GAN model generates textures conforming to the style of historical buildings. Depth maps are optimized through deep residual networks to generate semantic 3D models. Experiments on three types of historical buildings with dense brackets, structural beams, and garden decorations yielded average depth estimation error ranges ranging from 0.05 m to 0.11 m, and the overall accuracy of semantic segmentation can reach up to 96.12%. The style consistency of complex components remains above 90%, and the satisfaction of topological constraints exceeds 92.74%. This method provides certain technical support for high-fidelity semantic reconstruction of complex historical buildings.