Research on diffusion model image data completion technology for 3D reconstruction of cultural heritage

Archaeological documentation and long-term preservation of cultural heritage artifacts require the digital restoration and three-dimensional (3D) reconstruction of the objects. Nevertheless, surface degradation, cracks, erosion, and missing areas tend to result in missing part of the image data, which significantly decreases the accuracy of reconstruction and geometric consistency. This study introduces a hybrid framework of Conditional Denoising Diffusion Probabilistic Model (DDPM) which is an image completion method based on diffusion and a multi-view 3D reconstruction pipeline to produce high-quality reconstruction restoration results and accurate geometric recovery. The Cultural Heritage Synthetic Depth Dataset has been gathered into multi-view RGB images, depth maps, and camera pose information to be used in controlled experiments. In pre-processing, images are resized, normalized, augmented, and corrupted with artificial noise, which is controlled to create a structure mask and random dropout to mimic damage patterns in the real world. The distorted inputs are filtered through a Conditional U-Net-based diffusion model that gradually filters out noise and fills in missing areas whilst maintaining structural continuity and context. The multi-view images have been restored and then used in a Structure-from-Motion (SfM) step to estimate camera poses and create sparse point clouds, and then Multi-View Stereo (MVS) dense reconstruction and surface generation based on Poisson reconstruction and Marching Cubes to create complete 3D mesh models. Through experimental analysis, it is shown that the suggested method is superior to Autoencoder, Partial Convolution (PConv), and EdgeConnect baseline. The framework results in PSNR of 28.0 dB, SSIM of 0.90, and LPIPS of 0.133 in image completion, and Chamfer Distance (CD) of 0.0037, Hausdorff Distance (HD) of 0.0139, and F-score of 0.88 in 3D reconstruction accuracy, which substantiates improved structural fidelity and less propagation of geometric errors.