Multiple description coding using rotated permutation codes

Summary form only given. This paper proposes a technique that would address the problem of designing multiple description source codes for J channels. Instead of implementing optimal combining we propose to simply average the decoded output of the individual channels and then adjust the length of the resulting vector based on a theoretical analysis valid for permutation codes. The choice of using permutation codes comes from the fact that their low complexity makes high dimensional vector quantization possible, i.e. large TV's, and our simulations have indicated that the random generation of rotation matrices works well when the dimension is high. For low dimensions, different outcomes of the generated rotation matrices seem to yield quite different performance, meaning that the random design may not be as appropriate for this case. Hence, any vector quantization scheme able to perform quantization in high dimensions could potentially replace the permutation coding in the proposed scheme. We also extend the method to use a fraction rho of the rate R to quantize the quantization error of the decoded data, when all descriptors are received, rather then using the whole rate to quantize the individual descriptors. This improves the performance when receiving all the descriptors at the cost of a decreased performance when some of the descriptors are lost. Varying rho therefore produces different operation points in the tradeoff between side and central distortion. The main advantages of the proposed method are its relatively low complexity and its ability to easily implement any number of descriptions