Abstract
In this work, we propose a hybrid vector quantization scheme for general images that aims at exploiting similarities among detail signals in a multiresolution decomposition of the image. In a multiresolution decomposition, the image is decomposed into a set of subbands which are localized in scale, orientation and space. Our coding scheme consists of dividing each subimage of the multiresolution decomposition into square range blocks. The range blocks are matched against domain blocks chosen in the lower resolution subimage with the same orientation, and coded through a description of the map transforming the domain block into the range block. The pool of of domain blocks acts as a codebook for the range block, as in vector quantization, with the difference that the codebook is built from blocks inside the multiresolution decomposition. If the prediction procedure is not satisfactory with respect to a target quality, the block is coded using a geometric vector quantizer for Laplacian random variables.

This publication has 13 references indexed in Scilit: