Multiple filterbanks for image processing: Implementation issues

Abstract

Multiple filterbanks have recently been introduced as an effective tool for processing image data, in particular for the detection of directional singularities. The practical implementation of such a technique, however, encounters many intrinsic difficulties. Indeed, the fact that different, nontrivial and anisotropic scaling matrices are used at each multiresolution level can easily result in a highly inefficient way of carrying out the downsampling, upsampling and convolution operations involved in the process, and of recording and tracking their outputs. The aim of this paper is to propose a general strategy to overcome these implementation issues by properly reformulating the above operations and taking advantage of the representation of the scaling factors in terms of their associated Smith decompositions.