Research On Realistic Reconstruction Of Incomplete Three-Dimensional Geometric Complex Surfaces Based On GAN

With the vigorous development of scene construction,industrial inspection and reverse reconstruction of cultural relics,it is of great significance to improve the fidelity of digital models of three-dimensional objects or scenes.Realistic reconstruction of incomplete threedimensional geometric complex surfaces is the key content of digital reconstruction technology of target model in reverse engineering,which reconstructs damaged and incomplete areas according to the characteristics of existing data of digital model.The existing digital model reconstruction methods have some shortcomings,such as incomplete point cloud feature extraction,low fidelity of shape generation and unrealistic model.Therefore,the optimized encoder,GAN shape generation and CGAN texture generation are combined to solve the above problems.The main contents of this paper include point cloud feature extraction based on coding,and automatic restoration method of shape and texture information of cultural relics based on deep learning.Combining the characteristics of point cloud model of common objects and bowl-shaped cultural relics model data in the actual archaeological field,the shape and texture generation model of point cloud model is constructed,which provides technical guarantee for the geometric and texture reconstruction of incomplete threedimensional geometric complex surfaces.Aiming at the characteristics of uncertain density and irregular geometric structure in the local area of point cloud model,this paper proposes a local feature attention coding method based on geometric affine transformation to extract the geometric information of point cloud.Firstly,local points are transformed into normal distribution based on geometric affine transformation,which aims to describe the sparse and irregular geometric structure of point clouds and improve the characterization ability of local characteristics of point clouds.Secondly,the multi-head cross-attention and self-attention mechanism are introduced to encode the point cloud in combination with FFN,so as to perceive the geometric details of the point cloud model and implicitly model the local features of the point cloud,and reduce the influence of sparse and irregular geometric structures in local areas on the point cloud feature coding;Finally,the point cloud geometry is generated.Experiments show that,compared with the classical point cloud completion network VRC-Net,the point cloud completion accuracy in this paper is improved by 16.3%,and the feature extraction method can better improve the network’s perception ability of local shape features of point clouds,providing a better intermediate feature vector for the subsequent decoder structure.Aiming at the shortcomings of low fidelity and complex reconstruction method in the generation of point clouds in the missing parts of the model to be completed,this paper proposes a dual discriminant decoding method of missing shapes for fragment splicing model,which generates shape features from the features extracted by the encoder.The idea of the network decoder in this paper is to generate sparse sets of skeleton point clouds to obtain skeleton point clouds of missing point clouds,generate refined point clouds for the second time on the basis of skeleton point clouds,and further discriminate the refined point clouds generated for the second time,so as to improve the accuracy of generating shape information of point clouds.Experiments show that this network is more effective in the three-dimensional shape completion task of the bowl-shaped cultural relic model,and the average chamfer distance is 20.2%higher than that of the contrast network.At the same time,in the point cloud completion experiment of the public data set ShapeNet,it is shown that the average error of the proposed two-branch generating network in completing other shape features is smaller,and the effect is remarkable in the model completion task with simple shape information,which provides a foundation for the subsequent model realism and has stronger performance and good application value.Aiming at the problem that the lack of texture information in the missing parts of the shape completion model leads to the low fidelity of the model,this paper proposes a texture feature completion method of the three-dimensional model based on CGAN.Firstly,the scattered point cloud data are preprocessed;Secondly,by combining CGAN framework,the two-dimensional texture map of the three-dimensional model is deleted to generate a network;Finally,the generated two-dimensional texture information is mapped to the model to realize the texture mapping of the three-dimensional model.Experiments show that this network has the best effect in the texture generation task of the bowl cultural relic model,and experiments on public data sets show that this network has good generalization performance in the task of repairing and completing two-dimensional pictures,and realizes the automatic repair of texture information of three-dimensional models.Aiming at the technical requirements of model data acquisition and 3D modeling,this paper designs a set of software systems for data acquisition and 3D reconstruction of incomplete 3D geometrically complex surfaces based on the research of realistic reconstruction technology of incomplete 3D geometrically complex surfaces.Specific functions include:(1)Using SHINING SE scanner to obtain the digital model of the object to be completed.(2)The missing shape of the model to be completed is completed by using the point cloud completion network based on geometric affine transformation and local feature attention coding.(3)Using CGAN-based 3D model reconstruction texture feature completion method to reconstruct the texture of point cloud model.The work of this paper can provide theoretical and experimental basis for further research on digital modeling methods of threedimensional models such as cultural relics reconstruction,scene construction and face restoration.