Research On 3D Models Denoising Algorithm And Its Application In 3D Surface Reconstruction

Feature preserved mesh denoising is one of the most popular topics in the computer graphics. Most of the previous mesh denoising methods first use neighbor points to compute normal and curvature, then adjust the position of the vertexes. However, with the high intensity of noise, the normal and curvature are not accurate, it lead to the vertexes update towards the wrong direction. In this paper, we take the non-local self-similarity of the 3D models into consideration, and proposed a novel mesh denoising method which called space structure based group sparse representation. Furthermore, we extend this denoising method to point cloud based surface reconstruction.To fully use the character of the 3D model, the local space structure is constructed by collaborating both the vertexes and facets information, in this paper, which is called K-ring local space structure. In 3D model, the local space structure is unlike the regular patch which is easy to measure the distance between different patch in image. We propose adopt the iterative closest point(ICP) algorithm to define the similarity measurement between the two different space structure. By utilizing the non-local self-similarity of 3D model, we stack the similar space structure into a group, which is our denoising unit. Each group is further effectively represented with the group adaptive dictionary of local space structures. Accordingly, the denoising vertexes is represented by weighting the corresponding vertexes in all these groups. Experimental result proves that our proposed method achieves better performance than the other state-of-the-art mesh denoising methods, especially with the high intensity noise.Laterly, we extend the denoising algorithm to surface reconstruction which is based on point cloud. The point cloud is first denoising by space structure based group sparse representation method, but the point cloud only have the vertexes information, so, the local space structure is constructed by K-nearest neighbors. Then the ball-pivoting algorithm is used to the noise removed point cloud. The experimental result proves that our proposed surface reconstruction method is more effective and robust than other state-of-the-art method, especially with the noised point cloud.