Research On Partial Structure Extraction And Segmentation Of Terracotta Warrior 3D Model Based On Deep Learning Methods

In the field of cultural relic protection,3D digital restoration is one of the hottest research topics.The partial structure extraction of the 3D model of the terracotta warriors is of great value in the research of the terracotta warrior restoration and fragment splicing.At present,the traditional segmentation methods rely on the subjective experience of experts.The operation is relatively complex,whose process takes a long time,and the overall error is large.To solve the above problems,this thesis proposes two automatic segmentation methods for the 3D point cloud of the terracotta warriors to extract partial structure data.The research work of this thesis mainly includes:(1)To solve the problem that the traditional methods of segmentation and partial structure extraction of terracotta warrior point clouds by hand are time-consuming and the error is large,an unsupervised segmentation method for terracotta warrior point cloud based on the seed region growing method and graph convolution neural network is proposed,called SRG-Net.Firstly,to improve accuracy,this method estimates the normal value of each point with coordinates of the point cloud.This method obtains the neighborhood with the k-nearest neighbor algorithm,and then the point cloud is pre-segmented by the region growing method with the neighborhood results.Secondly,a neural network(SRG-Net)can learn the pre-segmentation results.Graph convolution can extract the point cloud features,and a multi-layer perceptron with dropout is used to segment the point cloud.The refinement method is used to optimize the final result.The experiment results show that SRG-Net can segment the terracotta warrior point cloud effectively and get high accuracy.Compared with the traditional manual method,SRG-Net can achieve higher accuracy without professional knowledge,which improves efficiency with less labor cost.Moreover,this method can get good results on the Shape Net dataset,showing its robustness.(2)To solve the problem that the former automatic segmentation method for the terracotta warrior point cloud is not sensitive to details,a few-shot segmentation method is proposed based on self-attention and fusion convolution neural network,called EGG-Net.Firstly,the spatial transformation network module is refined in order to improve overall accuracy and reduce latency.Secondly,the fusion convolution encoder is used based on graph and edge convolution.The graph convolution uses the k-nearest neighbor algorithm based on self-attention mechanism,which enables the neural network to learn local and global features.The edge convolution enables EGG-Net to learn the topology of the point cloud better.Secondly,EGG-Net uses the structure-aware loss function,which improves the training efficiency and accuracy of the experiment.Finally,fine-tune operation is added to the pipeline to optimize the experiment results.The experiment results show that EGG-Net can achieve better experiment results on the terracotta warrior dataset and Shape Net dataset.With a small amount of calibrated data,EGG-Net can get higher accuracy with less latency and memory used,so EGG-Net is more robust.