Mesh denoising with local guided normal filtering and non-local similarity

Most of existing mesh denoising schemes are inspired by the techniques of conventional 2D image denoising. However, due to the significant difference between 3D mesh and 2D image, the employment of some well-known natural image priors, such as non-local similarity, are not straightforward in mesh denoising. In this paper, we revisit natural priors in the context of mesh denoising, and propose an effective mesh denoising scheme by combining local normal smoothness and nonlocal self-similarity. Specifically, the normals of neighboring faces and the current face are weighted combined to suppress noise, according to the distances to the current face and their guidance normals. Furthermore, the normals of non-local faces with similar structures are exploited. To find similar structures and calculate the similarity, the concept of k-ring patch is introduced for each face, in which the consistency and average normals of patches are used for finding similar patches. At last, the distance and normal difference between faces are used in calculating similarity between patches and non-local normals will be weighted by similarity. Experimental results show that the proposed scheme outperforms the state-of-the-art techniques, in terms of both objective and perceptual metrics, especially for the meshes with regular structures.