Total-variation based picture reconstruction in multiple description image and video coding

摘要

This paper studies how to reconstruct pictures with the best possible quality upon receiving more than one description at the decoder side of a multiple description coding (MDC) system, assuming that an MDC encoder has been fixed at the encoder side to generate multiple descriptions. To this end, we formulate the problem into a total variation (TV) regularized optimization in which all received descriptions are regarded as targets to form multiple fidelity terms. Two solutions are then developed. First, we solve a standard Lagrange-type optimization involving multiple Lagrange multipliers, and this approach is applicable to any MDC encoder. Second, when multiple quantizers with different step-sizes or dead-zones are used to generate individual descriptions, we make use of the intersection of the overlapped quantization intervals (in the transform domain) in all received descriptions. Both solutions are demonstrated to offer a quality gain (subjective as well as objective) over what can be achieved in the existing methods. In particular, the second approach is found to offer the best gain consistently when a large number of descriptions are needed.