Mesh Denoising Based On Local Feature Matching

With the needs of film and television production,game entertainment,virtual envi-ronment,computer-aided medicine,the acquisition and processing of three-dimensional digital models' geometry is increasingly important.In recent years,the popularity of 3D scanners and depth cameras has greatly simplified the 3D geometric modeling process,making it easier for people to create 3D shapes from real-world objects.However,in the process of acquiring three-dimensional data,due to the unavoidable errors introduced by equipment and human,models obtained by scanning are often different from those artificially created by the artists,since they often contain various kinds of noise,and have certain deviations from the actual objects.The presence of noise greatly reduces the quality and visual effect of the 3D mesh,and has a significant impact on subse-quent applications,so it is an important to denoise the corrupted mesh.In geometric processing,removing noise in the scanned model has always been a classic problem.Mesh denoising is used to recover mesh from noisy data to obtain a high-quality model,and to remove the noise to reconstruct the mesh surface while maintaining the original topological and geometric features of the surface,and ensuring the recovered mesh will not shrink or require unreasonable manual processings and so on.There are extensive research works focusing on mesh denosing.Although mesh denoising has witnessed a great progress,it is still challenging in some aspects.Firstly,as sharp features and noise information are encoded as high-frequency signals,there is no good strategy to distinguish these two types.Secondly,many existing algorithms rely on the manually chosen domains of parameters in the algorithm,which are not automatic and limit the applications of these algorithms.Thirdly,most of the methods have certain presuppositions in the types of mesh and noise since they are varying in mesh denoising process,and therefore making these methods not applicable to general cases.In order to solve the problem of model denoising and improve some shortcomings of the existing methods,this thesis presents an algorithm based on local feature fitting.The algorithm is divided into two steps.Firstly,the normal of the mesh face is processed by the local regression function.And then a mesh with less noise is reconstructed ac-cording to the adjusted normal vector.The two steps are iterated until an ideal denoising result is obtained.The core idea of this thesis is to find the relationship between the noisy model and the original model by analyzing a large number of existing noisy models and ground-truth models.In this paper,we first use a local geometric feature description FND(filtered facet normal descriptor)to represent the local geometric and noise information of the mesh.Thereafter,we only need to consider the relationship between the local geomet-ric feature FND of noisy mesh and the facet normal of corresponding original mesh,and use this relation to get the calibrated normal vector of the noisy mesh.In the pre-processing stage,we explorer numerous existing noisy meshes and original real models to obtain corresponding pairs of the local geometric features FND in noisy mesh and the facet normal in original mesh.In the denoising stage,the local geometric features of the input nose mesh are first calculated.Secondly,the geometric features are used to match features stored in the pre-established database,and therefore the calibrated mesh surface normal vector is obtained.Finally,with calibrated surface vector information,the vertex positions are updated.Experiments show that our method can remove the noise and keep the sharp fea-tures of the mesh well,for models with artificial noise or scanned ones.In this thesis,the method is fully automatic and easy to denoise the noisy mesh.The method of this paper has no constraints on specific types of the mesh and the noise,and can be easily extended to various other data models.