Space-variant picture coding

Abstract

Space-variant picture coding techniques exploit the strong spatial non-uniformity of the human visual system in order to increase coding efficiency in terms of perceived quality per bit. This thesis extends space-variant coding research in two directions. The first of these directions is in foveated coding. Past foveated coding research has been dominated by the single-viewer, gaze-contingent scenario. However, for research into the multi-viewer and probability-based scenarios, this thesis presents a missing piece: an algorithm for computing an additive multi-viewer sensitivity function based on an established eye resolution model, and, from this, a blur map that is optimal in the sense of discarding frequencies in least-noticeable- rst order. Furthermore, for the application of a blur map, a novel algorithm is presented for the efficient computation of high-accuracy smoothly space-variant Gaussian blurring, using a specialised filter bank which approximates perfect space-variant Gaussian blurring to arbitrarily high accuracy and at greatly reduced cost compared to the brute force approach of employing a separate low-pass filter at each image location. The second direction is that of artifi cially increasing the depth-of- field of an image, an idea borrowed from photography with the advantage of allowing an image to be reduced in bitrate while retaining or increasing overall aesthetic quality. Two synthetic depth of field algorithms are presented herein, with the desirable properties of aiming to mimic occlusion eff ects as occur in natural blurring, and of handling any number of blurring and occlusion levels with the same level of computational complexity. The merits of this coding approach have been investigated by subjective experiments to compare it with single-viewer foveated image coding. The results found the depth-based preblurring to generally be significantly preferable to the same level of foveation blurring.