Cascaded Normal Filtering Neural Network for Geometry-Aware Mesh Denoising of Measurement Surfaces

Mesh denoising is a fundamental component of many disparate reverse engineering applications of measurement surfaces. This article presents a cascaded normal filtering neural network (termed a CNF-Net) for geometry-aware mesh denoising of measurement surfaces. CNF-Net leverages the geometry domain knowledge (GDK) that, a mesh approximates to its underlying surface compactly if all mesh facets at most lie on the surface intersections while not crossing them. Benefiting from the well-estimated underlying geometry of noisy mesh facets, a multiscale guidedly filtered normal descriptor (M-GFND) is formulated, and multiple height maps are constructed from the M-GFND. The height maps can be effectively fed into CNF-Net for learning the transformation matrices between the M-GFND and the ground-truth facet normal. CNF-Net can automatically handle meshes with multiscale geometric features yet corrupted by the noise of different distributions, while existing learning-based wisdoms commonly pursue an overall normal estimation accuracy yet fail to preserve surface significant features. Both visual and numerical evaluations on synthetic and real noise data sets consistently show the clear improvements of CNF-Net over the state-of-the-arts.

I. Introduction

Surface mesh data, which contain both point locations and their connectivities to discretely represent 3-D objects, are widely used in applications ranging from neural network training [1]–[3], industrial design [4]–[7], computer-aided medical detection [8], [9], to virtual clothes try-on [10], [11]. These meshes are commonly obtained by water-tightly reconstructing unstructured point clouds [12], [13], which are captured from various 3-D scanners [14] and depth cameras [15], or estimated from multiview photographs [16]. However, noise will invariably creep into the meshes, due to both measurement errors (e.g., vibrations or scattering of the 3-D scanner, and reflections, over- or under-exposure of the camera) [17] and computational errors caused by 3-D reconstruction and photogrammetry techniques [18]. To allow the most effective use of surface meshes in geometry processing, mesh denoising is the first essential task.