Research On Feature Extraction From Scattered Point Clouds And Reassembly Of Fragments With Defective Fractures

Using digital means to realize the virtual restoration of cultural relics,not only can effectively shorten the recovery cycle,but also can avoid the secondary destruction by artificial restoration.Therefore,the relevant researches have become a hotspot of computer graphics.The core of the virtual restoration of the cultural relics is the searching and reassembly of the adjacent fragments.The traditional geometry-driven techniques depend on the integrity and accuracy of the geometric information of the fracture areas.Thus,they fail to work on the fragments with defective fracture areas.Feature extraction is the cornerstone of the search and reassembly of the adjacent fragments.The accurate description of the shape of the fragments can only be obtained by effective features,bringing a great advantage for the subsequent searching and alignment of the adjacent fragments.Therefore,this dissertation focuses on the feature extraction methods of scattered point clouds,and applies the multi-feature fusion to the automatic reassembly of cultural relics fragments,in order to address the problem of automatic reassembly of defective fragments.The main research works and contributions are concluded as follows:(1)In traditional feature extraction methods,the local surface fitting depends on the prior knowledge,and the surface fitting quality directly affects the validity of the feature extraction.To solve this problem,an extraction method for valley-ridge features from point clouds based on local reconstruction was proposed.By constructing a triangular mesh that is close to the potential surface and can reflect the local geometry,on the basis of characteristic of differential "curl to straight",principal curvatures were calculated.Then according to the definition that valley-ridge points are the extreme points on the principal directions,the valley-ridge features were extracted based on the multi-scale ideology.The proposed method can effectively address the problem that the surface fitting results directly affect the validity of the feature extraction results.(2)The existing methods use global feature measurement threshold and the local geometric information to extract features,which result in sensitivity to the sharpness of features,and poor effects for models with different shapes of surfaces.Thus,a novel method for feature extraction from point clouds based on DBSCAN clustering was proposed.The reverse k nearest neighbors(kNN)of points were defined as a new feature detection operator.Then the definition of the global constraints of features was presented.Finally,the DBSCAN clustering method was utilized to cluster the points and extract features.The proposed method effectively solves the problem of feature sharpness sensitivity caused by using the global feature measure thresholds and only using local information of points,it outperforms the previous methods with respect to the models containing surfaces that have diverse geometries.(3)Aiming at the problem that the feature extraction method cannot adaptively calculate the feature measurement threshold,and avoiding the loss of potential shape of the feature points in the feature lines reconstruction,an approach for extraction of feature lines from point clouds using statistics was presented.By defining a new feature descriptor and introducing a new surface traversal model based on Poisson distribution,the proposed method marks the potential features;then the features were extracted by analyzing the region information of the potential feature points.Finally,the complete feature lines are reconstructed based on the geometry representations of the features that were created by adapting L1-median locally,instead of using distance/path parameters.The proposed method does not need any prior surface reconstruction and is not largely affected by noisy points,neighborhood scales or sampling quality.We demonstrate the benefits of our method with the favorable results for real-world point clouds with varying types of features.(4)To avoid forming holes in flat regions and loss of the shape of surfaces by existing point clouds simplification methods,a point clouds simplification method with geometric feature reservation was proposed.Based on the feature extraction methods in(3),the features and the average curving degrees of the potential surfaces are acquired.Then the shared nearest neighbor(SNN)clustering method was utilized to analyze the curving degrees of the local surfaces,and different simplification strategies were used to delete redundant points.The experimental results presented in this dissertation demonstrate that,the proposed simple and effective method can not only retain both the features and the curving degree of the potential surfaces,but decrease the simplification error as well.(5)In order to effectively address the problem that traditional geometry-driven methods fail to reassemble the fragments with incompleteness in fracture surfaces or break-curves,an automatic reassembly method based on skeleton graph matching was proposed,which transfers the matching of fragments into complementary matching of incomplete geometric texels.Firstly,the constraints for matching texels were designed according to the features of geometric textures.Then the skeleton graph of the incomplete texels was generated;and based on the skeleton graph grammar and the constraints,the incomplete texels were matched.Finally,the sequence of the incomplete texels was used as the upper constraint to search the adjacent fragments.Experimental results show that the proposed method can successfully solve the reassembly problem of the fragments with incompleteness in fractured surfaces and break-curves.It outperforms on the fragments with little local curvature changes of geometric texture.(6)In order to address the problem of automatic reassembly of defective fragments,especially for the fragments with large local curvature changes of geometric texture,an automatic reassembly method of the fractured objects based on the adaptive neighborhoods and multi-feature fusion was proposed.Based on the General Adaptive Neighborhood(GAN)and multi-feature fusion strategy,the proposed method transfers the matching of the contour curves to matching of the contour bands,so as to more accurately calculate the adjacency of fragments.Further,the Coarse-to-Fine strategy was also used to search adjacent fragments step by step,which can improve the searching effectiveness.The experimental results show that the algorithm can fuse various features on the fragment effectively,and outperforms the previous methods on the defective fragments.