Research On Mesh Denoising Technology

With the development of computer science,demand for high quality 3D models has appeared in many domains such as virtual reality,augmented reality,industrial design,film and game industry.Due to the difficulty of drawing huge models manually,scanning real objects is the most common way of obtaining high quality models.However,due to the accuracy limitations of scanning devices,raw mesh models are inevitably contaminated by noise.Hence,mesh denoising has become an active research topic in the area of computer graphic in recent year.The main target of mesh denoising is keeping feature while remove noise,which is as same as in image denoising.However,some operations are hard to be applied on the irregular edges and vertexes of meshes,such as finding similar structures,organizing data into regular form,which are common in image processing.What's more,there have been some algorithm designing for irregular data.However,most of these research focus on point cloud data.It is hard to process data that containing both edges and vertexes,and there are few relevant research.In view of the above questions,we propose three mesh denoising schemes in terms of overcoming the effect of irregular structure and designing algorithm for irregular structure.First,in this paper,a patch-based similarity enhanced face normal filtering scheme is proposed.The irregular structure of mesh improves the difficulty of finding and using similar structures.To solve this problem,in this paper the adjacent faces are composed into patches as the basic units for finding similar structures.Then the normal difference and distance between faces are employed to comput the similarity of patches.At last,three different schemes are proposed to exploit the usage of similar structure,and the comparison results between them and traditional schemes are shown.From the experiment results,the proposed schemes achieve better results than traditional schemes on objective results and show better feature recover results on models with small and large-scale features.Second,in this paper,a graph-based guided normal filtering scheme is proposed.Since meshes are composed by vertexes and edges,which can be easily transformed into graph.a graph-cut based guided normal filtering scheme is proposed in this paper.The proposed scheme contains two parts: First,the local faces are composed into patches,then each face is regarded as a vertex and undirected edges are built between adjacent faces.The normalized cut is applied on patches for extracting local features.Finally,the feature-preserving guidance normals are built according to the features,and the guided filtering is employed to finish denoising.What's more,some accelerating methods are proposed since the time cost of normalized cut is high.From the experiment results,the proposed scheme shows advantage on recovering features of models with complex structures.The accelerating method can save 99% computing time with negligible performance loss.Third,in this paper,a deep-learning based denoising schemes is proposed.Since it is hard for tradition networks to process irregular structures,the voxelization strategy is introduced to transform meshes in to volumetric representation as the input of networks.Then,a network structure and the corresponding training schemes are proposed,which can output the denoising normals according to the volumetric representation.From the experiment results,the proposed scheme shows great superiority in objective comparisons.What' s more,it introduces less pseudo-features and shows good feature recovering results on models generated by different scanners in subjective comparisons.All the three above algorithms have good robust and feature recovering results.They also achieve desirable results on both objective and subjective evaluations.What's more,three algorithms have their own advantages in some specific scenes.