基于权重引导的点云降噪及法向估计网络

Point Cloud Denoising and Normal Estimation Network Based on Weight Guidance

为了提高网络对局部邻域的特征提取能力,提升网络的降噪性能,现有的基于深度学习的点云降噪算法对每一邻域点计算一个权值刻画其与当前点位于同一几何结构的概率,通过筛选具有较大权值的点对邻域结构进行简化,提高网络特征提取的性能。但由于其未对权重学习进行有效的约束与引导,所学权重无法对邻域结构进行准确刻画。本文设计了权重引导,通过法向差异(法向权重引导)及欧几里得距离差异(距离权重引导)对邻域点与当前点位于同一几何结构的可能性进行预判,并将其用于约束邻域筛选网络的权重学习,提高权重学习和邻域筛选的准确性,进而提升整体点云降噪的质量。实验结果表明,本文的算法在降噪结果及法向估计方面均有提升,在不同噪声尺度下也更具鲁棒性。In order to improve the feature extraction ability of the network for local neighborhoods and improve the noise reduction performance of the network, the existing point cloud denoising algorithm based on deep learning calculates a weight for each neighborhood point to describe the probability that it is located in the same geometric structure as the current point, and simplifies the neighborhood structure by screening the points with larger weights to improve the performance of network feature extraction. However, due to the lack of effective constraints and guidance for weight learning, the learned weights cannot accurately describe the neighborhood structure. In this paper, we design a weight guidance to predict the probability that the neighborhood point is located in the same geometric structure as the current point through the normal difference (normal weight guidance) and the Euclidean distance difference (distance weight guidance), and uses it to constrain the weight learning of the neighborhood filtering network, so as to improve the accuracy of weight learning and neighbor-hood filtering, and then improve the quality of overall point cloud denoising. Experimental results show that the proposed algorithm is more robust in terms of noise reduction results and normal estimation, and is more robust at different noise scales.