Improved Multiple Description Framework Based on Successively Refinable Quantization and Uneven Erasure Protection

A method to produce balanced multiple descriptions (MD) of a source is by successively refinable quantization (SRQ) in conjunction with uneven erasure protection (UEP) [1], [2]. This work proposes an improvement to this balanced MD coding framework. In order to generate $L$ descriptions, the set of source samples is first partitioned into $L$ subsets of equal size, then each subset is quantized separately. Further, interleaved systematic Reed Solomon codes of codelength $L$ and decreasing strengths are applied across the streams output by the SRQs. The improvement over the previous UEP-based MD code is evaluated using the expected distortion of the source reconstruction at the decoder as a performance measure. For a Gaussian memoryless source, the asymptotical improvement in performance, as the rate and code block length approach $\infty$, can attain as much as 1.68 dB (for $L=3$ and very low probability of description loss), with a tendency to decrease as the number of descriptions and the rate of description loss increase. In the practical setting using scalar SRQ, small rates and small $L$, the observed improvement generally matches the asymptotical values.