Multistage index assignments for M-description coding

We consider the problem of generalized multiple description coding of image data, in which an image is encoded into M descriptions sent over M unreliable channels, and any description subset is useful in reconstructing an approximation of the original image. In a two-description scalar quantizer, a quantizer index is simply mapped by an injective mapping, defined by an assignment matrix, on a pair of indices. We propose a novel M-description coding method based on index assignment matrices for scalar quantizers. Instead of using a straightforward M-description generalization with hypercubes, we propose to encode a quantizer index in a multistage fashion, each stage doubling the number of descriptions. This method provides a much higher flexibility, allowing for instance different redundancy allocations for different loss profiles and simple repacketization.