Research And Application On Virtual Reassembly Technology Of Ceramic Cultural Relics Fragments Based On Global Similarity Of Fracture Surface

The pottery cultural relics unearthed by archeology have a long age and their materials are fragile.Due to crustal movement,weathering and erosion,improper preservation and treatment,the original intact cultural relics were damaged into several fragments.Cultural relic restoration technology can reproduce the original appearance of damaged cultural relics,which is of great significance to archaeological research and cultural heritage.Computer-aided virtual reassembly and restoration is an important technical solution in the field of cultural relics restoration,which has irreplaceable advantages compared to manual splicing restoration.The two key parts in the virtual reassembly method are the matching and registration process.The reliability of artificially designed feature descriptors used in the matching process of existing methods needs to be improved.In the registration process,the registration accuracy between fragments is the main goal.Based on the global similarity of fracture surfaces,this thesis takes improving feature reliability and registration precision of fragments as research direction and main contributions of research is as follows:(1)Aiming at the problem that the feature space constructed by the traditional artificially designed descriptor is relatively limited and the computational cost is tedious,this thesis uses deep feature descriptors,and designs a neural network called Siamese-Point Net which is based on Siamese Network and realizes the measurement of the global similarity between point cloud to evaluate the matching degree of fracture surfaces of fragments in the form of scores.Experiments show that this method achieves a high accuracy rate and provides strong support for the matching stage.(2)A coarse-to-fine registration strategy combining traditional methods and deep learning methods is proposed.Rough registration results are obtained through Principal Component Analysis(PCA);on this basis,Deep Closest Point(DCP)network after migration training is used to obtain detailed registration results.This method avoids the problems of iterative operation and non-converge,eliminates the influence of threshold,and improves the success rate and accuracy of registration to a certain extent.(3)A multi-fragment reassembly solution based on the global similarity of fracture surfaces is proposed.Aiming at the problem of reassembling multiple fragments in practical application scenarios,a complete processing solution based on a greedy algorithm is proposed.This solution is based on the matching score output by the Siamese network in(1),and fragments with the largest score are selected as the registration objects in each round;the registration method in(2)is used to obtain the reassembly result of multiple fragments,following the idea of greedy algorithm.