Multiple Description Image Coding Based on Multi-Stage Vector Quantization

A new design approach for multiple description coding, based on multi-stage vector quantizers, is presented. The design is not limited to systems with two descriptions, but is also well suited for the n-descriptions case. Inspired by the concept of channel optimized vector quantization, the design can easily be tailored to suit different erasure channels, e.g. packet erasure channels with memory (burst-losses). The optimization procedure used in the design takes a sample-iterative approach. All stage codebooks are updated simultaneously for each vector in the training database. The resulting algorithm has the behaviour of a simulated annealing algorithm, with several good properties, e.g. it usually provides codebooks with good index assignments. Image results are presented for systems with 2 and 4 channels. The image coder is based on a subband transform followed by 64-dimensional vector quantization, to illustrate the capacity of the design to handle large problem sizes.