Research On 3D Reconstruction Technology For Heritage Conservation And Digital Presentation

Cultural heritage,refers to people in the process of human historical development,rich in history,art,science and other research value of cultural relics and monuments.The research of 3D reconstruction technology for the conservation and digital presentation of cultural relics has important research significance and practical application value for the digital exhibition,restoration and conservation of cultural heritage.Among the various 3D reconstruction technology methods for cultural heritage,the multi-view 3D reconstruction method has high applicability,low reconstruction cost and good reconstruction accuracy.Combining with the current mainstream deep learning methods can further improve the accuracy and integrity of multi-view 3D reconstruction results,which is sufficient for the task of high-quality and high-fidelity3 D reconstruction of cultural heritage.This paper focuses on the application of deep learning to improve the reconstruction effect of multi-view 3D reconstruction methods,and proposes an end-to-end learnable adaptive depth interval multi-view stereo matching network for the problems of inaccurate depth estimation,incomplete reconstructed point cloud and low computational efficiency of current multi-view 3D reconstruction methods based on deep learning.Firstly,the multi-scale feature Transformer module is designed to extract features with different texture richness at different scales of the input image more efficiently.Then,a similarity metric is introduced to group the features to construct the matching cost between views,and redundant features are removed to improve the computational efficiency while ensuring accurate depth estimation.Finally,an adaptive depth interval module is designed based on a coarse-to-fine depth inference strategy to guide the depth refinement stage to delineate more accurate depth prediction intervals using the uncertainty of the coarser stage depth estimation.All estimated depth maps are then used to generate 3D point cloud models of the target objects by depth map filtering and fusion operations.The results of a series of comparative experiments and quantitative or qualitative analyses of this method on DTU benchmark,Blended MVS dataset and Tanks &Temples benchmark show that the multiscale feature Transformer module and adaptive depth interval module can effectively improve the accuracy and completeness of the reconstructed point clouds.Compared with other advanced methods,this paper can handle the multi-view 3D reconstruction of complex scenes more effectively,and has the advantages of low memory consumption and fast running speed.Finally,this paper designs the complete process of multi-view 3D reconstruction of cultural heritage,and applies the method to the high-quality 3D reconstruction of some cultural relics in Yunnan Provincial Museum,and the finalized 3D model of cultural relics can be further used for the conservation and digital presentation of cultural heritage.