Multi-resolution models for data processing: an experimental sensitivity analysis

Hierarchical Radial Basis Functions Networks (HRBF) have been recently introduced as a tool for adaptive multiscale image reconstruction from range data. They are based on local operation on the data and are able to give a sparse approximation. In this paper HRBF are reframed for the regular sampling case, and they are compared with Wavelet Decomposition. Results show that HRBF, thanks to their constructive approach to approximation, are much more tolerant to errors in the parameters when errors occurs in the configuration phase, while they are more sensitive to the errors which occurs since the network has been configured.