Multiscale Coding of Images

A video coding system is described which uses adaptive quantization of spatio-temporal subbands to enable transmission of near NTSC resolution color image sequences over a 1.5 megabit channel. The sequence is first decomposed into spatio-temporal spectral components using quadrature mirror filters. Simple heuristics are used to determine the local importance of a given component's energy. These components are then coded in one of three ways; either set to zero (coarsest quantization), vector quantized or scalar quantized. The representation of an image sequence as set of spatio-temporal coefficients is shown to offer great efficiencies to the task of vector quantization, specifically in minimizing the size of the codebook. Thesis Supervisor: Andy Lippman Title: Lecturer, Associate Director, Media Laboratory This work was supported in part by Columbia Pictures Entertainment Incorporated, Paramount Pictures Corporation and Warner Brothers Incorporated.