Joint source-channel decoding of variable-length encoded sources with applications to image transmission.

摘要

This dissertation presents one of the first solutions to the problem of designing optimal, maximum a posteriori probability (MAP) decoders for different variable-length encoded sources over channels without memory and the first such solution for channels with memory. Also, the complexity of the solutions presented in this dissertation remains a constant with time unlike that of the other solutions. Moreover, this dissertation presents the first application of the MAP-BSC algorithm to a subband based image coding system.; In the case of fixed length codes (with N codewords), the MAP problem reduces to processing log₂ (N) bits at one time. This is not a good solution when the codewords are of unequal lengths. When errors are introduced by the channel in a variable-length encoded signal, the partition of the bit stream into component codewords is no longer obvious. One of the main contributions of this dissertation is the introduction of the notion of complete and incomplete states in the state space associated with the dynamic programming formulation which lets us deal with the variable-length codewords in an elegant manner.; In the first part of this dissertation, we consider the design of the optimal MAP decoder for variable-length encoded sources over binary symmetric channels. We first consider Markov sources and then particularize it to memoryless sources. We also show how the proposed algorithm may be extended to Markov sources of higher orders. The second part of the dissertation proposes the design of the MAP decoder for variable-length encoded sources over channels with memory. Performance of both the decoders are compared to that of the standard decoders to demonstrate the superiority of the proposed decoders over the conventional ones. The robustness of the decoders under channel mismatch conditions is demonstrated experimentally. In the third part of the dissertation, the proposed decoder is tested on an image transmission system to show that the decoder performs significantly better than the conventional decoder both perceptually and objectively (in terms of the peak signal-to-noise ratio) even when the source is not “cleanly” modeled. Other salient features of the algorithms developed in this dissertation are that no assumptions are made on the number of samples or bits transmitted. (Abstract shortened by UMI.)