Terracotta army fragment splicing method and system based on fracture surface information

Abstract

The invention discloses a terracotta army fragment splicing method and system based on fracture surface information. The method comprises the steps of: dividing a terracotta army splicing data set into a test set and a training set; carrying out migration training on the point cloud registration network by using the training set to obtain a terracotta army splicing network model after parameter adjustment; selecting the registration direction corresponding to the minimum distance error as the optimal selection, and achieving optimization of the principal component analysis method; testing the optimized principal component analysis method by using the test set to obtain a rough splicing result of the terracotta army splicing network model; inputting the rough splicing result into the terracotta warrior splicing network after parameter adjustment to obtain a fine splicing result, and realizing terracotta warrior fragment splicing; and making a terracotta warrior terracotta data set to be used for achieving deep neural network migration training and testing, optimizing a traditional principal component analysis method, and adopting a distribution splicing strategy from coarse to fine, wherein the method is higher in splicing success rate and accuracy and lower in operation cost.

Description

technical field [0001] The invention belongs to the technical field of virtual restoration, and in particular relates to a splicing method and system for fragments of terracotta warriors and horses based on fracture surface information. Background technique [0002] As the funeral objects of Qin Shihuang, the terracotta warriors and horses have a history of more than 2,000 years, so many individuals broke into pieces. In order to save manpower and material resources, avoid duplication of labor and possible secondary damage, it is a mature idea to use computer technology for virtual stitching and repair. Different from the standard data set, the fractures of the terracotta warriors and horses are irregular, and the shape gap between the fragments is large. Therefore, in the process of stitching, the stitching results obtained by applying the existing 3D object stitching methods are not accurate enough. [0003] The Iterative Closest Point (ICP) method takes the distance betw...

AI Technical Summary

Problems solved by technology

Different from the standard data set, the fractures of the terracotta warriors and horses are irregular, and the shape gap between the fragments is large. Therefore, in the process of stitching, the stitching results obtained by applying the existing 3D object stitching methods are not accurate enough

[0003] The Iterative Closest Point (ICP) method takes the distance between two objects as the optimization goal, and gradually updates the correspondence between the points and the spatial position of the objects in an iterative manner. Large computational overhead, and easy to fall into the local optimal solution; the Normal Distribution Transform (NDT) method expresses the 3D point cloud as a rasterized normal distribution, and determines the matching point set through the matching of the normal distribution. The method is sensitive to parameters, and the success rate is not high; the Principal Components Analysis (PCA) method takes the three main direction coordinate axes of two objects as the processing object, and while reducing the amount of calculation, it also has certain defects:

[0004] (1) There is ambiguity in the registration direction between the main direction coordinate axes, and there are 4 different registration directions, which will lead to different results;

[0005] (2) Since the corresponding relationship at the point level is not calculated, the accuracy is not enough; existing deep learning methods such as DeepGMR, DCP, etc., the main problem lies in the data

[0006] Their training and testing processes use standard data sets, which have a large amount of data and the shapes of the objects to be spliced ​​are exactly the same. In reality, the data of terracotta warriors and horses are very different, and the data volume is relatively small. The accuracy of the model trained on the dataset directly applied to the splicing of terracotta warriors and horses will be low, and the data scale of the terracotta warriors and horses itself is not enough to support the training of new models

Images