A Log-Likelihood Ratio based Generalized Belief Propagation

摘要

In this paper, we propose a reduced complexity Generalized Belief Propagation (GBP) that propagates messages in Log-Likelihood Ratio (LLR) domain. The key novelties of the proposed LLR-GBP are: (i) reduced fixed point precision for messages instead of computational complex floating point format, (ii) operations performed in logarithm domain, thus eliminating the need for multiplications and divisions, (iii) usage of message ratios that leads to simple hard decision mechanisms. We demonstrated the validity of LLR-GBP on reconstruction of images passed through binary-input two-dimensional Gaussian channels with memory and affected by additive white Gaussian noise.