Research On Feature Extraction And Classification Of 3D Models Of Terracotta Warriors Fragments

Cultural relics are the most representative treasures of cultural heritage.Due to the low efficiency of traditional manual restoration and damage to cultural relics,it has become a necessary trend to use computers to digitally protect and restore cultural relics.It is mainly used for classification,matching,splicing and hole repair of cultural relic fragments.The number of cultural relic fragments is huge,the shape is irregular and incomplete,and the surface geometry and texture characteristics are complex.Therefore,this thesis uses the three-dimensional model of the terracotta warrior fragments as the research object to carry out the research on the classification of cultural relic fragments.The main research contents are as follows:1.Aiming at the problem that the feature extraction of the 3D model can only extract the local surface geometric features,and the depth information of the 3D model is lost,this thesis proposes a method of Slice-Center descriptors(SCs)as the digital feature of the 3D model Said.And based on the similarity measurement of the descriptor,it supports the realization of the classification and identification of the three-dimensional model.A threshold of 0.3 is set by the matching degree of the prior model category as the category evaluation standard to achieve the classification effect.Experiments show that when the threshold value of 0.3 is used as the classification judgment,the classification accuracy of the nail fragments of the terracotta warriors and horses fragments reaches 85.33%,which shows the feasibility of this method.2.For a large number of unknown fragments,the above methods are inefficient.Therefore,this thesis proposes a Point-Plane neural network architecture,which mainly focuses on the lack of input point cloud local connectivity information of the point cloud neural network Point Net,and then improves the Point Net network.This thesis adopts the improved method of "replacing points with faces" to form a point-face feature network.Experiments show that the average accuracy of the terracotta warriors fragments datasets is 82.11%,which is 3% higher than the average accuracy of Point Net,which shows the feasibility and effectiveness of this method for the terracotta warriors datasets.At the same time,the average accuracy rate for public data sets has reached 85.3%,which is 1% lower than that of Point Net,but it also shows the feasibility of this method.3.When constructing the local area in the sampling layer of the Point-Plane neural network,a method of "plain nearest neighbor sampling" is proposed,that is,iterative partial triangular patch "wrapping" method to construct the local area,which is more efficient than using iterative farthest point sampling.Because it only needs to select the point cloud contained around the center triangle surface,the number of calculations is relatively small,and the sampling efficiency is improved.