Mesh Denoising Based On Adaptive Filter

Weighted average is one of the most common strategies used in various state-of-the-art digital signal filters,and its performance depends on the weight design.When computing the weight between the current sample and one of its neighbours,existing methods consider only properties of the two samples,such as positions and values.They definitely tend to suffer from cross-region mixing.For example,assigning a large weight between two nearby samples separated by some feature edges,even when their properties are close,will damage the local structure.In this paper,we present a universal filter model for mesh denosing,such as the shortest path.It estimates the weight between the current sample and its neighbours based on the integral of two kinds of properties' difference along some path,connecting them.Therefore,prominent features are better preserved when removing textures or noises.It is very generalized and many classic filters are just special cases of it,such as geodesic filter and the propagated filter.Furthermore,a specialized intrinsic filter based on the generalized model is introduced for mesh denoising by integrating along geodesic paths.Besides,considering that even geometrically complex shapes can be characterized by a rather small number of features,which means large normal differences are sparse,the L1 norm is employed in the integral.On the other hand,we develop the intrinsic filter with different parameters for modeling the geometry features and model a regression function,we could obtain a denoise filter which is used for mesh denoise in the same environment.We follow the common two-stage mesh denoising framework,the position of vertices are updated after the face normals are filtered,and the two steps are iterated.Experiments illustrate the enhanced efficacy of our intrinsic filter comparing with state-of-the-art methods.