Research On Surface Reconstruction Based On Feature Points

During 3D reconstruction of medical images,based on image segmentation algorithms,a serialized CT image will generate several triangle mesh of different organs and some of these organs,such as bones,consist of numerous vertices and triangle patches.A powerful server is required to run segmentation algorithms and store tremendous organ models.But when querying mesh model,it is not necessary to transmit a whole model,which consumes a lot bandwidth and delays users.In this paper,a method to dynamically transmit 3D mesh model was discussed.With this method,the server transmits point cloud data with feature points while the client uses these data for surface reconstruction.At last,corresponding evaluation algorithms and application scenarios were introduced.The approach in this paper mainly consists of the following parts:Part I discusses the detection algorithms of feature points for triangle mesh.Based on spatial topological characteristics or spectrum domain characteristics of triangle patches,corresponding operators(calculation formula)were applied by these algorithms to calculate the significance value of each vertex.According to the significance value of each vertex,local and global feature points are calculated.Related detection algorithms of feature points were discussed and modified in an attempt to detect enough feature points for surface reconstruction.The detection algorithms of feature points studied here were categorized to feature point algorithm based on spatial neighborhood,such as Harris-3D,and feature point algorithm based on spectrum domain,such as Stochastic Mesh Laplacian.In experiments,these feature point algorithms were applied to medical 3D model and standard 3D model for detection and comparison.Part II describes surface reconstruction algorithms of point cloud,mainly including Power Crust,SSDRecon and Screened Poisson.Reconstruction results and performance differences were compared in experiments.It was found that the Screened Poisson has better comprehensive performance than another two algorithms and was applied to the surface reconstruction of point cloud algorithms with feature points.The point cloud model presented here was made from feature points detected by the feature point algorithm and random points collected by Poisson disk sampling algorithm.Random point algorithm is used to maintain the whole topology shape of the model,while feature point algorithm is used to maintain details of important areas.In addition,differences of two vertex normal algorithms were illustrated.Part III mainly presents relevant evaluation algorithms,the indicator SSIM is used to evaluate the similarity in 2D vision of reconstructed model and original model,while discretescale axis is adopted to evaluate the similarity in 3D shape and structure.Experiments show that the mesh model reconstructed by the point cloud model is highly similar to the original model in 2D vision as well as 3D shape and structure,and that a better option is to set the vertices of the point cloud model to 30% of those of the original model during reconstruction.At last,what mentioned is the combined application scenarios of surface reconstruction algorithm,mesh retrieval algorithm,and mesh segmentation algorithm,which are mainly employed to dynamically transmit 3D model to reduce the data to be transmitted.