Multichannel 2-D power spectral estimation and applications.

Abstract

Spectral estimation for multiple 2-D signals by model-based methods is developed. The procedures compute the entire spectral matrix of autospectra and cross spectra for the set of 2-D signals. Spectral analysis by autoregressive (AR) modeling is studied extensively. Specific differences between AR models for this problem and those for lower dimensional problems are highlighted. An extension of the Jackson-Chien method for combining estimates with single quadrant support is proposed and a method is developed for estimating the model parameters directly from the data (i.e. without prior estimation of a correlation matrix). A measure of the similarity of two spectral estimates based on the statistical divergence is proposed and used to compare various spectral estimates. A comprehensive set of experimental studies are presented showing the performance of the methods in estimating the autospectra and magnitude and phase of the cross spectra. The Maximum Likelihood Method (MLM) of spectral estimation is extended to the multichannel 2-D case. The properties are compared experimentally with the autoregressive methods. The Improved Maximum Likelihood Method (IMLM) is also developed for the multichannel case. Finally applications of multichannel 2-D spectral analysis models to image coding are presented.