Deep Learning Based Efficient Mesh Parameterization

Mesh parameterization is a fundamental and very important problem in computer graphics and digital geometry processing.Many problems related to digital geometry processing,such as texture mapping,compression deformation,editing,mesh optimiza-tion and mesh subdivision,require parametric transformation as an important part of the problem solving.The problem of mesh plane parameterization can be regarded as a mapping problem between mesh plane and plane parameter domain.In recent years,the research on mesh parameterization has made extensive progress and many parameter-ization methods have appeared.However,the existing mesh parameterization methods based on optimization are slow and not robust enough.At the same time,with the de-velopment of deep learning methods,the convolutional neural network has obtained the most advanced results among a large number of problems in computer graphics and computer vision,and the convolutional neural network based on graph data has been put forward one after another,which brings a new idea to the mesh parameterization method.In this paper,we apply the idea of deep neural network to the grid parameter-ization problem,and build a graph convolutional neural network model for a specific type of mesh to obtain the parameterization results of a specific type of mesh.Main tasks include:In graph convolution mesh encoder automatically,on the basis of building up a kind of graph convolution neural network model.Graph Convolutional neural network can maintain the topological connection of the input data in the output,which provides help for grid parameterization with deep learning.In order to support mesh data types,using a fast localization of convolution filter as the convolution kernels in graph con-volution neural network,and use the Chebyshev polynomial convolution kernels do K order approximation,to achieve the purpose of fast calculation,at the same time us-ing multistage convolution kernels to improve the convolution kernel feature extraction ability.The above-mesh sampling layer and the under-mesh sampling layer based on the mesh data type are established to replace the pooling layer corresponding to the convolutional neural network,expand the sensing field,and obtain the global charac-teristics of the grid data in the network.The deep network is constructed by alternating convolution layer and sampling layer,and the deep network ensures the extraction of data features with stronger representational ability.In terms of network optimization.Dropout method and regularization method are adopted to alleviate the problem of net-work overfitting and improve network performance.And establish a number of network model comparison,determine the final network model.Finally,the established model of a convolutional network of parameterized graphs for a specific grid type is used to ob-tain corresponding grid parameterization results for different parameterized distortion measurement functions.Finally,by comparing with the traditional algorithm,it is verified that while ensuring the parameterization effect,the established graph convolutional network model can be used to obtain the parameterization result faster,Among them,the speed gap is obvious in the mesh data of large size.And the face mesh can be processed in batches,and multiple parameterization results can be obtained at the same time.