Salih Ergut

Optimized Equiripple-Based Decorrelating Filter Banks

Wednesday, August 30, 2000
1:00 PM
206 Egan

Abstract

The use of filter banks in adaptive/statistical signal processing applications can result in reduced complexity and faster convergence rate. In particular, binary-tree-structured filter banks provide further reduction in the complexity of adaptive filtering via the use of a sparse multirate configuration. However unless the subband channels are perfectly decorrelated, the use of a sparse adaptive filter results in performance degradation (i.e., increase in the mean square error). In order to keep this degradation to a minimum it is essential to optimize the frequency response of the filter bank to achieve the lowest possible inter-channel correlation.

While QMF filter banks are popular in speech and image processing applications because of their perfect reconstruction property, their subband channel decorrelation capability is far from optimal. In contrast, this research adapts the classical Parks-McClellan equiripple filter design technique to binary-tree structured filter banks, thus obtaining a flexible trade-off between subband decorrelation and inter-channel transition bandwidth. In this sense it outperforms the QMF approach, which offers a much more limited trade-off between channel decorrelation and transition bandwidth.

Thesis Committee:
Prof. H. Lev-Ari (Theses advisor) Prof. M. Schetzen
Prof. B. Shafai