A Parametric Modeling Approach to Image Compression

This work examines a new approach for lossy image compression based on parametric modeling. The basic concept seeks to exploit the high resolution and energy compaction properties of auto-regressive (AR) spectrum estimation given a limited number of parameters on a sparse set of high amplitude waveforms. In its simplest implementation the compression scheme treats each image row as a 1-D spectrum. A symmetric extension of the image row is used to insure that the transformed data remains real- valued. An inverse Fourier transform converts each row image into a waveform that can be modeled using standard linear prediction techniques. The image compression parameters therefore become the reflection coefficients commonly used for speech compression. Further compression can be achieved by exploiting the correlation between the rows of AR coefficients. There are many ways to do this. In this paper the method used was based on a simple decimation/interpolation of the AR parameters of the rows of the image.