Research On 3D Mesh Denoising Algorithm

3D sensing and scanning technology has been widely used to capture the digital surface of physical objects.However,in the process of capturing and 3D reconstruction,noise from various internal and external factors inevitably enters,which reduces the quality and usability of surface data and obviously hinders subsequent geometric processing tasks.Mesh denoising is an important tool in geometric processing and has many applications in computer graphics,such as computer aided design,reverse engineering,virtual reality and medical diagnosis.The main goal of mesh denoising is to remove noise while maintaining clear features.Although some traditional methods can effectively remove most of the noise,they will more or less blur the features and lose details at the same time.Therefore,we need an effective feature-preserving mesh denoising algorithm.In recent years,the following challenges still exist in the mesh denoising task: Firstly,the best parameters of most of the most advanced methods are usually manually selected and remain unchanged in the whole model and the whole denoising process.Secondly,how to prevent the mesh volume from shrinking in the process of mesh denoising;Third,how to effectively combine the classic task of mesh denoising with the latest machine learning to make the denoising model real-time and effective.Aiming at the above challenges,this paper proposes a targeted and effective mesh denoising algorithm.For the first two problems,we propose a two-step iterative algorithm in which most of the parameters are adaptive and automatically adjusted,and the vertex update of the algorithm can prevent volume shrinkage and flip triangle to a certain extent.The effectiveness of the method lies in: 1)In each iteration,the effect of preserving features is achieved according to the different adaptive adjustment parameters of different region feature intensities in the mesh.2)As the iteration progresses,the smooth term is gradually reduced by automatically adjusting the parameters along with the change of iteration times,thus achieving the function of preserving detail features at the later stage of iteration.3)In vertex update,constraint items are introduced to pull the vertices back to the previous position to prevent the mesh volume from shrinking.In view of the third problem,we propose a data-driven mesh denoising method based on recurrent neural network,which learns the relationship between the feature descriptor of the mesh surface and the ground-truth normal.Using the self-feedback characteristic of the recurrent neural network,the output of the hidden layer is not only related to the current input,but also related to the output at the previous time,thus the mapping function from the feature descriptor of the face to the normal of the face can be more accurately established.Finally,the position of the updated vertex is adjusted according to the estimated face normal.A large number of visual and numerical results show that our two-step iterative method is very effective compared with the prior art methods,and the model proposed in our data-driven method can obtain better real-time denoising results on various models and noise types.