Robust mesh denoising based on collaborative filters

抄録

Mesh denoising is crucial for improving imperfect mesh models corrupted with raw or synthetic noise. The main technical challenge of mesh denoising is to faithfully retain geometric features when removing noise. Nevertheless, most of the existing approaches are not suitable for large-scale noise, overlook some geometric features, and often need carefully select appropriate parameter values to achieve a good result. In this paper, we present effective collaborative filters for mesh denoising based on vertex classification. The effectiveness of our method stems from several aspects: 1) Global preprocessing is performed to drastically reduce noise influences. 2) Normal tensor voting is utilized to classify the vertices, so that we can select different filters to estimate the face normals according to the detected type of vertices. Our method adaptively prevents the side effects from facets with high geometrical disparity in the feature region, which avoids the subjective selection of parameter values to achieve the local optimum, and removes outliers in the non-feature region. 3) Normal difference weights are introduced to vertex updating. Benefited from the well-designed filters on different types of vertices, our algorithm produces visually and numerically better denoising results than the existing typical approaches for both CAD and generic models corrupted by high level of noises, especially at sharp features, such as edges and corners