A model for landscape heritage information extraction and 3D generation from sketches and text prompts with IoT-enhanced data fusion
Yu Xing, Yu Xing, Hanlin Chen, Yimeng Xing, Yimeng Xing · Alexandria Engineering Journal · 2025
The advancement of deep learning methods has spurred research on generating 3D digital landscape models from sketches and text. However, the scarcity of high-quality reference data and the absence of precise geometric data pose significant challenges, impeding existing methods from achieving a balance between geometric accuracy and texture fidelity in the digital reconstruction of landscape cultural heritage. To address these problems, this paper proposes a sketch-and-text prompt-based landscape heritage information extraction and 3D generation model (LHIE-3D), that integrates sketch-guided geometry generation, text-driven texture refinement, and multi-view consistency optimization to produce high-fidelity 3D reconstructions. Evaluation on the ShapeNet-Sketch3D and TU-Berlin Sketch datasets demonstrates that LHIE-3D achieves a CLIP similarity score of 0.81, significantly outperforming ImageDream (0.55) and Sketch2Model (0.60). Subjective evaluation results indicate that LHIE-3D attains scores of 4.63 for text consistency, 4.36 for sketch fidelity, and 4.35 for unedited region preservation. Experimental results confirm that LHIE-3D not only enhances reconstruction accuracy but also offers intuitive editing capabilities, providing a practical tool for cultural heritage preservation.