Multi-resolutive sparse approximations of d-dimensional data

• Authors: Giuseppe Patanè
• Venue: Computer Vision and Image Understanding
• Year: 2013

This paper proposes an iterative computation of sparse representations of functions defined on R^d, which exploits a formulation of the sparsification problem equivalent to Support Vector Machine and based on Tikhonov regularization. Through this equivalent formulation, the sparsification reduces to an approximation problem with a Tikhonov regularizer, which selects the null coefficients of the resulting approximation. The proposed multi-resolutive sparsification achieves a different resolution in the approximation of the input data through a hierarchy of nested approximation spaces. The idea behind our approach is to combine a smooth and strictly convex approximation of the l"1-norm with Tikhonov regularization and iterative solvers of linear/non-linear equations. Firstly, the iterative sparsification scheme is introduced in a Reproducing Kernel Hilbert Space with respect to its native norm. Then, the sparsification is generalized to arbitrary function spaces using the least-squares norm and radial basis functions. Finally, the discrete sparsification is derived using the eigendecomposition and the spectral properties of sparse matrices; in this case, the computational cost is O(nlogn), with n number of input points. Assuming that the data is supported on a (d-1)-dimensional manifold, we derive a variant of the sparsification scheme that guarantees the smoothness of the solution in the ambient and intrinsic space by using spectral graph theory and manifold learning techniques. Finally, we discuss the multi-resolutive approximation of d-dimensional data such as signals, images, and 3D shapes.