Research On Feature Extraction And Semi-Supervised Classification Algorithm Of Terra Cotta Warriors Fragments

The Qin Terracotta Warriors is a powerful proof of the brilliant civilization of the Chinese nation,but most of them were broken when they were unearthed,so it is of great significance for their splicing and restoration.Traditional manual restoration requires expert experience,is time-consuming and laborious,and it is easy to cause secondary damage to cultural relics.There is an urgent need for computer-assisted virtual restoration to provide guidance schemes.However,during the splicing process,there are too many pieces of Terracotta Warriors fragments,and most of them are seriously damaged and disorderly.The direct splicing workload is too large.If they can be classified and then restored,it will greatly reduce the calculation workload and improve the accuracy.However,due to the fact that there are too few labeled samples of fragments,it is necessary to study the semi-supervised classification technology.Therefore,in view of the characteristics of large scale,high latitude,and few labeled samples of Terracotta Warriors fragments,this thesis studies anchor-graph-based semi-supervised classification method and auto-encoder-based feature extraction method,and designs and implements Terracotta Warriors fragments feature extraction and semi-supervised labeling algorithm.The main work of this thesis includes:(1)Aiming at the limitation of anchor point selection in the anchor graph semi-supervised classification algorithm,the ADPC-AGR algorithm is proposed.Use the adaptive density peak clustering algorithm to obtain the anchor point,explore the spatial structure of the data by calculating the local density and relative distance of the data,find the maximum point of relative density,and select the top k points with the highest relative density as the clustering center,can obtain more representative anchor points.The experimental results show that the ADPC-AGR algorithm is superior to the existing anchor graph algorithm in classification accuracy and calculation efficiency.(2)Concerning the dimensional disaster problem for data such as images,using the excellent feature learning ability of convolutional neural network,a convolutional sparse auto-encoder is designed as a feature extractor.Automatically learn the high-level features of the image through multi-layer convolution,extracts the output of the bottleneck layer as effective features,and designs a weighted loss function to optimize the performance of the network.Combining the convolutional auto-encoder and the semi-supervised classification algorithm,a Terracotta Warriors fragments feature extraction and semi-supervised labeling algorithm is designed.The experimental results show that the convolutional sparse auto-encoder has a good feature learning ability.In the semi-supervised classification experiment of Terracotta Warriors fragments,only 10% of the labeled samples achieved a classification accuracy of 93.2%.(3)A set of semi-supervised classification system for automatic labeling of Terracotta Warriors fragments is designed and developed,integrating the relevant algorithms in this thesis,which can realize the large-scale Terracotta Warriors fragment labeling task and improve the automation of Terracotta Warriors fragments virtual restoration degree.