Research On 3D Mesh Models Denoising Algorithm

Due to the rapid development of 3D scanning technology,we can get high-precision 3D models,however,due to various factors,such as limited device accuracy,messy surrounding environment,etc.,the acquired raw data inevitably contains noise.In order to make better use of these 3D models in further geometric processing,it is necessary to denoise these noisy 3D models.This paper presents two algorithms to denoise the mesh model.The first is a traditional feature-preserving mesh denoising framework,the second algorithm uses deep learning to denoise the mesh model.The main challenge in mesh denoising is to remove noise while maximally preserving sharp features.Therefore,this paper presents a mesh denoising framework based on the guided normal.Firstly we extract the features by dividing the faces into feature faces and non-feature faces.Then for the two types of faces,we use joint bilateral filtering by different neighborhoods to filter face normals respectively.In particular,for feature faces,we use the similar metric we defined to produce partial neighborhood in which these faces have similar geometrical feature to feature face.Lastly,we update vertex positions according to the filtered face normals.Experiments on a large number of models demonstrate the effectiveness of the proposed denoising framework,and compared with the existing methods,the framework we designed can better restore the obvious geometric details.Deep learning has achieved great success in 2D images in the last few years,so we want to apply deep learning into the mesh model,but the existing neural network framework can’t directly use the mesh model as the input.Therefore,this paper designs an algorithm to remove the mesh noise by convolutional neural network.Firstly,the normals of faces in the mesh are updated by convolutional neural network,and then the vertex positions are adjusted to match the updated face normals.In this algorithm we propose a method to map the face normal in mesh model into 2D images,and design a convolutional neural network architecture that updates the face normals.Finally,the experimental results show that the denoising algorithm proposed by deep learning can effectively remove the noise in the mesh model.