Three-dimensional non-separable subband decomposition for coding of time-varying images

Abstract

Proposes a three-dimensional (3D) spatio-temporal subband decomposition of time-varying imagery by means of 3D non-separable filters. A subband partitioning of the image content is sought such that it exploits the distribution of the 3D spectrum of time-varying imagery for image coding and transmission. Within a subband, a discrimination of the directional distribution of the localized spectrum due to motion is considered. The authors examine a critically sub-sampled subband filter bank system with the property of near-perfect reconstruction. Non-separable 3D filter and subsampling structures are chosen for the analysis to exploit the spectral directionality. A class of QMF filter bank systems is explored with angular frequency space discriminating properties. In particular, a 3D QMF prototype is considered possessing the spectral-domain rhombic-dodecahedral symmetry. Results for selected image sequences are presented.