Research On Semi-Supervised Collaborative Classification Of Terra-Cotta Warriors Fragments Based On Graph

The Terracotta Warriors are the typical representatives of ancient civilization in China.However,due to the long-term influence of natural factors and the destruction of manual excavation,few intact Terracotta Warriors have been unearthed.Therefore,it is urgent to piece together the excavated fragments and restore them.Classification of fragments can greatly improve the efficiency of splicing.However,in the existing fragments data,the amount of labeled data is very small,while the number of unlabeled data is large.In view of this situation,semi-supervised classification technology can ful y explore the helpful information carried in unlabeled data and improve the classification accuracy.Among many semi-supervised classification algorithms,the graph-based algorithm is the most active and widely used,and the co-training algorithm can fully consider the multi-view features of data.Therefore,based on these two algorithms,this thesis carried out relevant research on the classification of Terracotta fragments.The main research contents are as follows: First,a multi-view representation method is proposed for the complex relationship data of Terracotta Warriors.Since the bipartite graph can only reflect the data association relationship between views,and the k-nearest graph can only reflect the similarity relationship between data within views,these two graphs have certain limitations.Therefore,this thesis combines the bipartite graph and the k-nearest graph,and make some improvements,propose a multi-view representation method for the fragments data of Terracotta Warriors: by establishing multiple graph structures at the same time,the similarity relationship of fragments data within a single view and the association relationship of fragments data between multiple views are reflected in the graph structures,and the graphbased semi-supervised classification algorithm is applied to each view feature of fragments data.Second,in order to avoid the problem that the overall classification performance is degraded due to the large performance difference between multi-classifier in collaborative training,this thesis proposes a multi-classifier collaborative training algorithm with smoothness and consistency evaluation indexes.Due to the serious degradation of the attributes and characteristics of the fragments,the multi-classifier learned on the multi-view features can not achieve the ideal classification effect.To solve this problem,the multiple classifiers obtained in this thesis are collaboratively trained.In the process,the smoothness and consistency evaluation indexes are joined.Aimed to take advantage of the differences between multi-view features,optimize complementarity reasonably,and help improve the classification performance and generalization ability of each classifier,at the same time,smoothness can make a single classifier have the same prediction ability for labeled and unlabeled data,and consistency can make multi-classifier predict the same label for unlabeled data as far as possible,so as to avoid the overall classification performance is degraded.Third,the design and implementation of the Terracotta Warriors fragments classification system.The semi-supervised classification algorithms involved in this paper is applied to the classification of terracotta warriors fragments to realize the functions of image preprocessing,feature extraction and classification of the parts.