Research And Application Of Denoising And Simplifying Methods Of 3D Cultural Relics Point Cloud

Digital collection technology forms a comprehensive and real permanent digital file for cultural relics,so as to realize the sharing of cultural relics information resources.It is a necessary technical means to preserve historical heritage and cultural heritage.With the improvement of computer software and hardware technology,the laser scanning technology of non-contact measurement is widely used in the collection of cultural relics data.Laser scanning technology has the advantages of fast speed and high accuracy in obtaining point cloud model,but the collected point cloud data of cultural relics are often affected by the reflection of surrounding objects,illumination,human factors,and even the inherent noise of the acquisition equipment itself,including a large number of redundant points,repetitive points and noises.The noise data in the point cloud of cultural relics will directly affect the quality of post-registration,segmentation and reconstruction of the point cloud of cultural relics.A large number of redundant data points in the point cloud of cultural relics will occupy a large amount of computer storage space,which will seriously affect the performance of computer modeling and rendering,and reduce the work efficiency and interactive experience of processing and displaying the later point cloud of cultural relics.Therefore,it is necessary to denoise and simplify the point cloud of cultural relics on the premise of ensuring the original texture characteristics of the point cloud of cultural relics.The research work of this paper is as follows:(1)To solve the problem that unsupervised point cloud denoising network can’t effectively remove outliers,a point cloud denoising network based on neighborhood estimation of deep neural network is proposed.Firstly,through the first network layer,the global features and local features are fused,and the outlier point set in the noise point cloud is estimated by paired outlier point labels.Then,remove that estimated outlier noise point set in the noise point cloud;Finally,the noise point cloud as a whole passes through the second network layer,and the location of the corresponding clean point is estimated by the neighborhood of the noise point,and the spatial prior term is introduced to guide the noise point cloud to converge to the manifold mode closest to the surface of the clean point cloud,so as to realize point cloud denoising.Experiments show that the algorithm can effectively remove outlier noise points and keep their texture features for large-scale noisy cultural relic point clouds with complex features.(2)Aiming at the deficiency of(1),which may lead to slight deformation of denoised point cloud,a point cloud denoising network based on displacement estimation of deep neural network is proposed.Firstly,through the first network layer,the global features and local features are fused,and the outlier point set in the noise point cloud is estimated by paired outlier point labels.Then,remove that estimated outlier noise point set in the noise point cloud;Finally,a local patch is established for the noise point cloud,and through the second network layer,the noise point cloud is guided to converge to the clean point cloud surface by correcting the displacement vector of the noise point cloud,so that the point cloud can be denoised.The experimental results show that the point cloud denoising network has excellent denoising performance,and the ability to keep sharp features of the point cloud model is outstanding,both for ordinary point clouds with different noise intensities and large-scale cultural relics point cloud models with natural noise points and complex features.(3)Aiming at the problems that traditional point cloud simplification algorithm is easy to delete feature points by mistake,resulting in holes,loss of boundary points and large reconstruction error in the simplified point cloud model,a point cloud simplification method constrained by curvature features based on knn is proposed.Firstly,the topological structure of point cloud is constructed by kd tree.Then,the knn algorithm is implemented to search the k nearest neighbors.Then,estimate the normal vector.Finally,by comparing the angle between the normal vectors and the threshold of each point,the points are divided into characteristic points and non-characteristic points,which are simplified by bounding boxes with different side lengths.The experimental results show that the algorithm can effectively keep the original details and geometric features of the point clouds and avoid holes on the premise of ensuring the simplification rate,whether it is a small-scale common point cloud or a large-scale cultural relic point cloud with complex geometric features.