Traditional digital twin geometric scene construction is often inefficient due to manual modeling, while existing automated solutions require expensive equipment and complex operations. To overcome these limitations, this paper proposes a low-cost and automated digital twin scene generation method based on image understanding. Monocular images from industrial production lines are processed using YOLOv11 for equipment recognition and Depth-anything-v2 for depth estimation to obtain initial spatial layouts. A dual-column layout optimization algorithm is further designed to correct monocular depth distortion by combining DBSCAN-based clustering and principal direction alignment with spacing constraints, resulting in standardized and coherent industrial layouts. In addition, a semantic 3D model asset library is built, and a large language model is utilized to perform semantic matching based on equipment functions and process attributes, enabling automatic retrieval of the most appropriate digital twin models. Experiments show that, for typical dual-column industrial layouts, the proposed method improves construction efficiency by 90.28% compared with manual modeling while achieving 90.5% matching accuracy. The prototype system confirms that the method effectively reduces development cost and enhances automation, providing a practical solution for scalable deployment of digital twin technology in industrial environments.
Chen et al. (Mon,) studied this question.