Neural networks for coefficient prediction in wavelet image coders

Abstract

We present a unique method for estimating the upper frequency band coefficients solely from the low frequency information in a subband multiresolution decomposition. First, a Bayesian classifier predicts the significance or insignificance of the high frequency coefficients. A neural network then estimates the sign and magnitude of the visually significant information. This prediction model allows us to construct an image coder which can exclude transmission of the upper subbands and reconstruct this information at the decoder. We demonstrate results for a two level subband decomposition.