Partial differences as tools for filtering data on graphs

High-dimensional feature spaces are often corrupted by noise. This is problematic for the processing of manifolds and data sets since most of reference methods (and especially graph-based ones) are sensitive to noise. This paper presents pre-processing methods for manifold denoising and simplification based on discrete analogues of continuous regularization and mathematical morphology. The proposed filtering methods provide a general discrete framework for the filtering of manifolds and data with p-Laplacian regularization and mathematical morphology. With our proposals, one obtains filters that can operate on any high-dimensional unorganized multivariate data. Experiments will show that the proposed approaches are efficient to denoise manifolds and data, to project initial noisy data onto a submanifold, and to ease dimensionality reduction, clustering and classification.