Research On Refinement Classificaiton Of Cultural Relic Fragments And The Method Of Reassembling For Multiple Fragments

Virtual restoration of cultural relics has been a hot issue in the field of heritage protection in recent years.Using computer science to reassemble the fragments virtually can help avoiding secondary destruction by human and speed up the restoration process.However,the fragments come with a large variety of geometric shapes due to damage,as a result.computer-assisted virtual restoration for cultural relics still faces the following challenges:First,it is inevitable that the digitalization process would bring in noises due to the limitation of environment and scanner.Important information,for example,surface finishing of the 3D model may be smoothed out during denoising because both noises and sharp features are high-frequency information.Second,as there is no refined classification of numerous cultural relic fragments,the existing automatic reassembling methods encounter many difficulties such as the complex adjacency relationship of the fragments and the high time complexity of reassembling.Third,the damaged surfaces of cultural relic fragments resuls in the loss of certain geometric features,which leads to the failure of reassembling or incorrect reassembling of the fragments.This paper focuses on the three core technologies in the virtual restoration process of cultural relics: the removal of noises on the surface of the 3D model,the refinement classification of cultural relic fragments and the reassembling of multiple cultural relic fragments.The significance of the dissertation are concluded as follows:(1)A point cloud denoising method based on Markov-Graph Laplacian regularization for the cultural relic is proposed.With this method,the 3D point cloud is divided into blocks based on the self-similarity theory of the block,and a graph model with Markov property is created.The graph laplacian regularization(GLR)prior is employed and the problem of point cloud denoising is formulated via maximum a posteriori(MAP)estimation to find out the most probable noiseless point cloud.Experimental results show that the proposed denoising algorithm results in more outstanding performance than competing methods in terms of visual effect,mean square error(MSE)and signal-to-noise ratio(SNR).(2)A refinement classification method of 3D cultural relic fragments based on shape feature extraction is proposed.The fragments are finely divided into specific parts according to the body features of a relic and then reassembled together.The process can greatly improve the efficiency of virtual restoration of cultural relics.Based on scale-invariant heat kernel signature,a low-dimensional shape descriptor is constructed with the bag-of-word approach,and then an unsupervised classification algorithm based on clustering is developed for classification.The experimental results show that the proposed descriptor can,to some extent,accurately recognize the shape features of 3D models.For models with distinct shape features the accuracy is more than 90%.This method has not only the high classification accuracy,but also the advantage of unsupervised learning.It also meets the needs of cultural relic samples whose category labels are not marked or cannot be marked.(3)A classification method of 3D cultural relic fragments based on deep learning is proposed.A deep hierarchically-connected network is designed based on the adversity-generation strategy,and dynamically generates enhanced samples that are more conducive to classification in the continuous training,which effectively solves the problem of lacking samples with the existing end-to-end deep hierarchically-connected network.Dynamic parameters are introduced in loss function of the data enhancement network to control the difference of cross-entropy loss between the enhanced sample and the real sample,so that the enhanced sample,though different from the real one,has similar shape characteristics with the real one.The focus loss function is employed to design the classification network.It makes the network focus more on the small samples and the samples that are difficult to classify in the training process,thus effectively reduces the negative effects due to the unbalanced sample size.Experimental results of public datasets show that the proposed method has higher classification accuracy than the competing methods.The experimental results of the real datasets show that the proposed method can significantly improve the classification accuracy of the samples that are difficult to classify.(4)A multiple fragment reassembling method based on key point descriptor is proposed.The well-preserved cultural relic as a template to guide the reassembling of fragments,the key points of the original surface of the template and fragment are detected,and the adjacent relationship between multiple fragments by calculating is determined by comparing the pieces of key point descriptors,to avoid the exhaustive search for adjacent fragments.At the stage of accurate matching of fracture surfaces,the feature points of the fracture surface and the curves along the principal directions of feature point clusters are taken as matching features simultaneously,so that the fragments can be effectively matched even when the amount of features is not sufficient.The method combines the shape features of the original surface and the geometric features of the fracture surface,achieving higher precision than the competing methods.