Synthesis filter bank with low memory requirements for image subband coding

One of the factors that makes subband coding seem less attractive for image processing compared to transform coding, is the extensive memory requirements for 2-D filtering. In this paper we present a filtering method that significantly lowers the memory requirements. A parallel, uniform and separable filter bank with 8 channels in each dimension is considered. It is shown that by a data reordering and a minor structural change to the filter bank, a blockwise filtering is made possible, thus reducing the total memory to 192 B registers and 9 KB RAM for an image of size 1152/spl times/1920 pixels.