3D Model Reconstruction and Texture Detail Restoration Technology of Digital Sculptures Using Masked Autoencoder

To protect and share cultural heritage, it is important to accurately recreate three-dimensional (3D) digital sculptures. But photogrammetry- and CNN-based approaches that are already out there typically have problems with occlusions, partial scans, and reflecting surfaces, which lead to geometric errors and texture loss. To tackle these issues, we provide a single Masked Autoencoder (MAE)-based architecture that can rebuild geometry and restore high-fidelity textures at the same time. The architecture combines RGB, depth, and fluorescence scans with a hybrid loss function that balances Chamfer Distance, perceptual similarity, and structural consistency. Fluorescence-guided texture enhancement makes it much easier to find small pigment differences and microstructural features that are typically missed in regular RGB imaging. We confirmed the suggested strategy using a freshly created Digital Sculpture Dataset (DSD) that includes 120 high-resolution cultural objects. Quantitative assessments exhibited enhanced performance compared to leading-edge benchmarks, including PointNet++, AtlasNet, and NeRF versions, attaining reduced Chamfer Distance, elevated F-scores, and superior perceptual metrics (PSNR, SSIM, and LPIPS). Qualitative evaluations validated the precise reconstruction of both macroscopic geometry and intricate surface textures, even under scenarios of occlusion or deterioration. The suggested framework is a strong and flexible way to preserve digital information with high accuracy. It may be used in virtual museums, conservation monitoring, and heritage research. This study improves digital archiving by combining geometry reconstruction and texture restoration into a single end-to-end process. It also opens up new possibilities for historically accurate 3D modeling of cultural relics.