A feature-preserving framework for point cloud denoising. (October 2020)

• Record Type: Journal Article
• Authors: Liu, Zheng; Xiao, Xiaowen; Zhong, Saishang; Wang, Weina; Li, Yanlei; Zhang, Ling; Xie, Zhong
• Is Part Of: Computer aided design. Volume 127(2020)
• Publication Date: 2020-10
• Subjects: Point cloud denoising -- Feature preserving -- Second order regularization -- Tensor voting -- RANSAC; Computer-aided design -- Periodicals; Engineering design -- Data processing -- Periodicals; Computer graphics -- Periodicals; Conception technique -- Informatique -- Périodiques; Infographie -- Périodiques; Computer graphics; Engineering design -- Data processing; Periodicals; Electronic journals; 620.00420285
• DOI: 10.1016/j.cad.2020.102857
• Languages: English
• ISSNs: 0010-4485

Abstract

Abstract: Point cloud denoising has been an attractive problem in geometry processing. The main challenge is to eliminate noise while preserving different levels of features and preventing unnatural effects (such as over-sharpened artifacts on smoothly curved faces and cross artifacts on sharp edges). In this paper, we propose a novel feature-preserving framework to achieve these goals. Firstly, we newly define some discrete operators on point clouds, which can be used to construct a second order regularization for restoring a point normal field. Then, based on the filtered normals, we perform a feature detection step by a bi-tensor voting scheme. As will be seen, it is robust against noise and can locate underlying geometric features accurately. Finally, we reposition points with a multi-normal strategy by using a simple yet effective RANSAC-based algorithm. Intensive experimental results show that the proposed method performs favorably compared to other state-of-the-art approaches. Graphical abstract: Highlights: An anisotropic second order regularization method is presented to restore the point normal field. A bi-tensor voting scheme, combining the normal and point tensor voting, is proposed to detect features on the noisy input. A simple yet effective RANSAC-based algorithm is introduced to estimate the multiple normals at each feature point.