A Signal-Specific QMF Bank Design Technique Using Karhunen-Loéve Transform Approximation

Block Wavelet Transforms (BWTs) are orthogonal matrix transforms that can be obtained from orthogonal subband filter banks. They were initially generated to produce matrix transforms which may carry nice properties inheriting from wavelets, as alternatives to DCT and similar matrix transforms. Although the construction methodology of BWT is clear, the reverse operation was not researched. In certain cases, a desirable matrix transform can be generated from available data using the Karhunen-Loéve transform (KLT). It is, therefore, of interest to develop a subband decomposition filter bank that leads to this particular KLT as its BWT. In this work, this dual problem is considered as a design attempt for the filter bank, hence the wavelets. The filters of the decomposition are obtained through lattice parameterization by minimizing the error between the KLT and the BWTmatrices. The efficiency of the filters is measured according to the coding gains obtained after the subband decomposition and the experimental results are compared with Daubechies-2 and Daubechies-4 filter banks. It is shown that higher coding gains are obtained as the number of stages in the subband decomposition is increased.