| In recent years,with the popularity of 3D laser scanners and low-cost sensors,the acquisition of 3D point cloud data has become faster and easier.The powerful representation capability of point clouds makes them widely used in various engineering fields.However,due to the equipment and environment,the scanned point cloud data inevitably contains a large number of noisy points.Therefore,as a fundamental step in point cloud geometry processing,point cloud denoising becomes crucial.Point cloud denoising has made great progress in the past decades.However,most of the current methods mainly target the ideal synthetic point cloud noise,or have unrealistic assumptions,such as ”assuming local surface smoothing” and ”assuming average noise distribution”.In order to solve these problems,this paper proposes a structure-aware and combined denoising method.The main work and innovation points of this paper are as follows.(1)A structure-aware denoising method is proposed.The method performs denoising by learning prior knowledge of external and internal point clouds while preserving the structural information of noisy point clouds.In this paper,non-local self-similar patches are first grouped from a set of external clean point clouds.Then,a Gaussian mixture model learning algorithm is used to learn the external non-local self-similar prior on the group of surface patches.This paper next learns internal priors from a given noisy point cloud in the same way to improve the prior model.The external prior may not fit the input noisy point cloud,while the prior learned from a given noisy point cloud may be inaccurate due to the interference noise.Therefore,in this paper,the denoised normals are estimated by integrating external and internal information through a set of hybrid orthogonal dictionaries.After obtaining the denoised normals,a feature-aware point update method is proposed in this paper.The method relocates the points by performing adaptive neighborhood selection for feature and non-feature points separately to match the obtained normals.Finally,this paper conducts a large number of experiments on both synthetic and real noise,and the experiments show that the method in this paper achieves good comprehensive performance in both objective and subjective evaluations compared with many state-of-the-art methods.(2)A boosting denoising framework based on integrated learning is proposed.This framework can denoise noisy point clouds with different levels and types of noise at the same time by cascading multiple denoising units.In this paper,we first implement an encoding-decoding based deep denoising unit.This denoising unit takes surface patches as input and encodes them into a 1024-dimensional vector after a multi-layer MLP,and then decodes them through a fully connected layer.The output of each denoising unit and the original point cloud are used as the input of the next unit,and the final denoising is achieved by multiple cascades.The experimental results on the mixed level and mixed type of noise data and real point cloud data proposed in this paper show that the method in this paper can achieve better results on most of the models. |
