Research On 3D Construction Of Nuo Masks Based On Deep Learning

As an ancient primitive culture,Nuo culture is an important part of Chinese traditional culture and has also been included in the national intangible cultural heritage list.Due to improper preservation of the Nuo masks,most masks were damaged to some extent or even disappeared.In the end,a single 3D Nuo mask retained only part of the Nuo mask pictures or a single Nuo mask picture.Therefore,using Nuo mask images to obtain their corresponding 3D Nuo mask shapes is very important for the protection of Nuo culture.In recent years,deep learning has shown its own development potential in many fields,and successfully applied deep neural networks to the 3D construction of single images.In this paper,the Nuo mask of the Maonan is selected for related research,and a 3D construction method of a single Nuo RGB image based on deep learning is designed.The main work is as follows:(1)Establishment of Nuo Masks 3D Database of the MaonanThis article uses the Observer 3D scanner to scan the Nuo mask of the Maonan nationality museum to obtain preliminary 3D data.Then,we used Geomagic Studio 12 software and ZBrush software to complete the preliminary repair and fine repair of Nuo masks 3D data,respectively.Based on the correction of Nuo mask world coordinate system and object coordinate system by Blender software,the corresponding 3D model is derived and sampled to obtain a high-low resolution 3D voxel model.(2)Establishment of Nuo Masks 2D image set of the MaonanThis article renders the produced 3D model with different rotation angles,random lighting,and a fixed color background,and crops the center of the rendered 2D image to obtain an RGB Nuo image with the same size.Then,this paper obtains the high-low resolution 6-axis orthogonal 2D depth map of each 3D Nuo voxel model to complete the production of Nuo depth image set.(3)Design of low-resolution 3D construction voxel networkBased on the traditional convolutional neural network,a multi-feature fusion lowresolution Nuo voxel 3D construction model is proposed.First,Multi-feature fusion method is used to enhance the learning ability of the network to extract the features of different layers of neural network in the stage of 2D encoder.Then,an extended 3D sub-pixel convolutional decoding algorithm is used in the 3D decoder stage.The algorithm can reduce model decoding time without loss of reconstruction performance.On this basis,a quadratic optimized low-resolution 3D Nuo network algorithm can be used to reconstruct more complete 3D Nuo masks.(4)Design of high-resolution 3D construction voxel networkA high-resolution 3D voxel construction model based on multi-residual dense block and a high-resolution 3D voxel construction model based on generative adversarial networks are designed respectively.Based on the generated low-resolution 3D Nuo mask voxels,the corresponding 6-axis depth map is obtained and fed into the depth map superresolution network,and the generated super-resolution depth image is used to complete high resolution 3D Nuo mask construction.Finally,the experimental results show the effectiveness of the method designed in this paper on the high-resolution 3D Nuo voxel construction task.Through the research on the above aspects,this paper completed the 3D construction task of a single Nuo mask image of the Maonan.Solved some problems in the production of Nuo masks related datasets and model design,and achieved good modeling results.