Statistical model-aided decoding of continuous-valued syndromes for source coding with side information

The concept of syndrome plays undoubtedly a central role in distributed source coding. With known source-side correlation, systems based on continuous-valued syndromes have indeed been shown to perform close to the Wyner-Ziv bound, both in theory and in practice. This paper investigates the application of the continuous-valued syndrome-based approach to the real case, where little or no knowledge regarding the source-side correlation is available at the encoder. Since in this case the encoder cannot operate at his best, traditional maximum likelihood decoding do not perform well. Iterative, factor graph based, statistical model-aided decoding is instead able to provide more accurate results. The experiments show in particular that model-aided decoding leads to about one order of magnitude less reconstruction errors within a few decoder iterations, which amounts to an increase of the signal-to-noise ratio of up to 3 dB.