Research On Hole Restoration And Fracture Surface Based Splicing Of Cultural Relic Fragments

Over 310 items(sets)of terracotta figurines,carriages and horses,and other various cultural relics have been unearthed from the third excavation of the Terracotta Warriors.However,almost all of these relics were fragments.Facing such large amount of fragments excavated,it is simply impossible to restore their original appearance manually.Although the modern three-dimensional scanning technology could be employed to obtain the virtual restoration of those fragments,there are still four problems of the technology as follows:1.There is no effective way to express the disordered 3D point cloud model,which enables the point cloud heritage model to directly carry out deep learning and other automated,intelligent operations;2.Due to the limitation of scanning environment and scanner,there are holes in the fragments of cultural relics obtained by laser scanner,and the model is incomplete;3.The extraction of matched fracture surface relies on manual labeling or feature-driven,and the automatic recognition of fracture surface driven by model data is not realized;4.The existing fragment registration algorithms has been designed feature descriptors based on knowledge of personal experience domain,and rely on traditional geometric features to complete fragment matching,that do not realize automatic feature extraction and fragment registration.In addressing the problems mentioned above,this dissertation intends to give deeper expression and analysis of point cloud data,then the fragments holes obtained by scanning are repaired,and the fragments are spliced by extracting the fractures.The major search tasks of this dissertation include:(1)A point cloud heritage data representation method based on the octree and 3D KD tree is proposed.This method firstly transforms the unorganized point cloud into a voxel space,and uses the octree to divide 3d model and improved octree coding method;Secondly,the adjacency relationship between nodes is constructed,and the octree is constructed in parallel on the GPU;Finally,a 3D K-D tree is used to index a single 3D spatial point in order to overcome the inefficiency of octree coding retrieval.Experiments showed that the method can truly reflect the details of the model itself,accelerate the construction of point cloud model and improve retrieval efficiency of point cloud model.Achieve the point cloud is transformed into the new data structure that convolution neural network can receive.(2)A repair method of complex holes in relic fragments based on the variational level sets proposed.Turn the problem of extracting the hole edge into a problem of spherical evolution of a hidden surface,and complete the hole repair using convolution and synthesis.Experiments show that the algorithm can effectively restore and maintain the detailed features of the holes.The algorithm solves the self-intersection phenomenon of complex hole repair well,and can repair island hole better.The algorithm is not only suitable for point cloud data,but also can be used to repair holes in the large-scale point cloud models.It achieves the goal of high quality,effective,complete and non-self-intersecting restoration of complex holes in the 3Dl model of cultural relics fragments.(3)An extracting method for the fracture surface of cultural relics fragments based on semantic segmentation is proposed.It established an end-to-end,data-driven multi-layer CNN(convolution Neural Network).First,the point cloud is converted into conventional input form that can be received by CNN for training.Then,the network learns the training data automatically,focusing on the contextual labeling feature of the cultural relics.Secondly,the full connection condition random field is added to the network to further refine the segmentation.Finally,the fracture surface feature is predicated and marked by classifier.Experiments show that the deep learning network can learn the features of the data set without manual intervention.When evaluating with test data,the algorithm achieves remarkable improvement in segmentation accuracy compared with the existing algorithm.The segmentation accuracy reaches about 89%,which is higher than the current mainstream method of fragmentation of cultural relics,and at the same time realizes automatic semantic marking.(4)A method for registration of cultural relics fragments based on feature extraction is proposed.In order to provide multi-feature matching for point cloud model,the convolution neural network is used to learn the global and local features of the fracture surface on the data set.First,use the data set to train the feature extraction ability of the convolution neural network to extract the key registration points,and label the key points waiting for registration in the global features by the classifier.Then study independently the local feature of each optimal neighborhood through the convolution neural network,and find the corresponding relationship of the two fragments waiting for registration according to the local features of the key points.The matching method solves the problem of the large differences but registration occurred in the global features.The experimental results showed that learning global features is more effective than the manual labeling or adaptive methods in obtaining the key registration points;the registration point pairs obtained by the registration based on the local feature extraction are more effective and accurate.By eliminating outliers and weighted adjacency matrix,the initial value of rough registration is accurate enough,the registration errors have been significantly reduced compared with the existing registration methods.