Research On Feature Extraction And Classification Algorithm Of Cultural Relic Fragments

With the development of computer image processing technology and visualization technology in recent years,cultural relics and its fragments can be digitally protected and restored by computers,which can improve the work efficiency of cultural relic fragment restoration and promote the integrated development of computer science and archaeology.Due to the huge amount of cultural relic fragments,complex geometric texture structure and incomplete information on the fracture surface of the original texture of cultural relic fragments,if the fragment is not classified before the restoration of cultural relics,the complexity of searching and matching adjacent fragments during the restoration of cultural relics will be increased,thereby reducing the overall efficiency of the restoration process of cultural relics.For this problem,an effective method is to classify fragments of cultural relics to narrow the search space of adjacent fragments.This article takes Terracotta Warriors and Horses of Qin Shihuang Mausoleum as the research object,and carries out research on the related technology of fragment classification.The specific research contents are as follows:1.On account of the limitations of a single global feature descriptor or local feature descriptor,a local neighborhood information and significant multi-feature descriptor is proposed.Firstly,the unevenness is defined based on the multi-scale volume integral invariant of the model surface,and the calculation results are processed by K-means clustering algorithm to obtain the model's distinctive feature point clusters;then,the significant connected point set is defined and the centroid point of the point set is selected as the feature point;finally,a local reference frame(LRF)is established locally at the calibrated feature points,and the local surface of the point cloud is continuously rotated and projected.The two-dimensional images are performed by feature encoding to complete the feature description and representation.This descriptor has scale invariance in describing both the local neighborhood features and the surface significant features of the fragment model,and also has strong robustness in describing the relics fragments with irregular size,irregular shape,complex surface and complex spatial structure.2.Based on the segmentation algorithm of the three-dimensional model,a template library of features of terracotta warrior fragments is established.A rough matching method based on volume integral invariant is proposed to calculate Gaussian curvature,and completed by multiple matching constraints that are established based on the geometric characteristics of the fragments.Then,the point pair characteristic distance of local neighborhood information and significant multi-feature descriptor is used to calculate similarity and judge the classification.According to experimental results,this algorithm,with high classification accuracy in fragment classification,is suitable for classification situations that a library of fragment templates has been created,the number of fragments to be classified is too few to be used for deep learning,and the structure of fragments to be classified is relatively complete.3.The geometric features of the point cloud model are lost due to the defect of cultural relic fragments,and the traditional method is easy to fail.A branching multi-feature network classification depth network based on the combination of 2D image and 3D point cloud is proposed and used for the first time to classify terracotta warrior fragments.First of all,according to the characteristics of terracotta warrior fragments,the data is pre-processed by uniform points selection and multi-view two-dimensional image projection;then,MVCNN and Point Net point cloud deep learning network structure are improved,and image information and point cloud information are fused by building a multi-branch structure,making up the shortcomings of the point cloud feature by the two-dimensional image features and the feature loss in the multi-view projection process of the two-dimensional image by the point cloud features,which become complementary features.The experimental results show that the classification accuracy of this algorithm can reach 85.11% on the data set of the terracotta warriors.This algorithm is suitable for classification situations that the number of fragments to be classified is high and the structure of the fragments to be classified is relatively incomplete.