Deep Learning Based 3D Reconstruction Of Ceramic Artworks

Ceramic culture is an integral part of the excellent traditional Chinese culture,and ceramic artworks are the crystallization of ceramic culture and China’s thousand years of porcelain smelting skills,which are the embodiment of artistry and practicality and have great appreciation value.However,due to the fragility of ceramics,it hinders the spread of ceramic culture.With the development of 5G technology and VR/AR display technology,technology and art in one ceramic art appreciation also increasingly penetrate into people’s lives.However,the lack of digital model resources of ceramic artworks also seriously restricts the process of ceramic culture promotion and dissemination with ceramic artworks as the carrier.Research on 3D reconstruction of ceramic artworks can solve the problem of lack of digital model resources of ceramic artworks,and create a virtual display of ceramic artworks in the form of 3D models in 5G platform under the premise of reducing costs,time and saving resources,which is a useful initiative to promote the appreciation of ceramic artworks into the life of ordinary people.This thesis carries out research and practice of 3D reconstruction of ceramic artworks based on deep learning by studying the 3D reconstruction algorithm of ceramic artworks based on deep learning and proposing an APMNet 3D reconstruction algorithm combining deep learning and traditional algorithms,the main research contents and work are as follows.1.Research on ceramic artwork dataset: In this paper,by analyzing the problems of reflections,repeated textures and no textures in traditional 3D reconstruction of ceramic artworks,the first dataset of diverse ceramic artworks was collected using dual cameras and structured light scanning,and 120 different ceramics were photographed in 7 different lightings from 49 precise camera positions,while precise structured light scanning was performed.A total of 32340 images were collected,forming a dataset of 30870 images for training and 1470 images for testing,providing data support for the experiments in this paper.2.3D reconstruction algorithm based on DLPMNet: This paper proposes to combine deep learning with traditional 3D reconstruction algorithm,and use deep learning to overcome the defects of traditional algorithm while inheriting the accuracy of traditional algorithm.Combining hierarchical independent prediction and multi-layer fusion prediction,forming a multi-scale feature extraction as the main framework,and fusing3 DPMNet algorithm to build the DLPMNet 3D reconstruction algorithm from coarse to fine.The experimental results show that DLPMNet is better than Colmap and Open MVS in the completeness of point cloud reconstruction and better than Colmap in the overall quality,and the texture details（legs,tail,bullhorn）reconstructed by DLPMNet are more complete,which solves the problems of no texture and reflection of traditional algorithm to some extent.3.APMNet-based 3D reconstruction algorithm for ceramic artwork: This paper proposes to optimize the DLPMNet algorithm using adaptive Patch Match.For the problems arising from the 3D reconstruction algorithm based on deep learning,on the basis of DLPMNet,it retains its multi-scale main framework from coarse to fine,Patch Match propagation,removes the Patch Match normal vector operation with high memory occupation and long running time,applies the low-resolution layer to triangulate the initial depth map,and uses local perturbation,adaptive propagation,adaptive spatial cost aggregation and loss function to optimize the DLPMNet-based 3D reconstruction algorithm and construct a dynamic APMNet 3D reconstruction algorithm.The results show that APMNet reduces 70.313% and 82.857% compared with Cas MVSNet,55.017% and 81.053% compared with UCSNet,70.513% and 85.484%compared with CVP-MVSNet,and 10.925% and 40% compared with DLPMNet in terms of memory consumption and running time and 40%;meanwhile,APMNet reduces 19.059%and 40.714% in accuracy difference and overall error compared to MVSNet,14.844% and21.513% compared to R-MVSNet,7.887% and 17.207% compared to DLPMNet,and5.143% in overall error compared to CVP-MVSNet.APMNet not only achieves a significant reduction in memory usage and time consumption generated by deep learning,but also further reduces the accuracy difference and overall error of model reconstruction.