Image coding with residual quantisation using statistically-selected quantisers

摘要

A pixellated image encoding method comprises: obtaining a residual image (difference between original and predicted image portions); encoding the residual by applying first probabilistic coding mode(s) to residual blocks segmenting the residual image according to a first segmentation, each residual block associated with a block type, wherein the first probabilistic coding modes each implement a quantization of a residual block using first quantizers selected (chosen, from a pool) based on first statistics on the residual image; obtaining a residual error image from the residual and the encoded residual image; encoding the residual error image, by applying second probabilistic coding mode(s) to residual error blocks segmenting the residual error image according to a second segmentation, wherein the second probabilistic coding modes each implement a quantization of a residual error block using second quantizers selected based on second statistics; and inferring the second statistics and/or the second segmentation from the first statistics, the first segmentation and/or the encoded residual image. Optimal quantisers may be selected, based on quantization mean error. Generalised Gaussian Distributions may be used. The method (and associated devices) may be applied to hierarchical scalable video coding (SVC), using base (BL) and enhancement (EL) layers.