A Two-Stage Estimation Approach for Bridging the Performance-Complexity Gap in Joint Source-Channel Decoding

摘要

Joint source-channel decoding techniques capitalize on residual redundancy that typically remains following a source encoding operation. These methods, which include MAP and MMSE-based decoders, estimate the sequence of encoded source symbols based on statistical knowledge of both the channel and the encoded source. Generally, these techniques are based on a Markov model for the quantized source and, thus, on a hidden Markov model for the source-channel tandem. The number of states in the hidden Markov model, and thus the computational and storage complexities, grow exponentially with the order (if) of the Markov model, i.e., the complexity order is O(N~(K+1)T), with N the number of source quantization levels and T the length of the data sequence. Thus, to retain implementable complexity, low order models (K = 1,2) are typically used, at the expense of model accuracy. In this work, we propose a method to bridge the performance-complexity gap, i.e. to provide solutions that give better performance than a low order decoder while incurring only modest increases in complexity. Our decoding approach consists of two stages: 1) low order JSC decoding, followed by 2) a linear FIR filtering of the JSC decoded signal. The linear filter is chosen to provide an optimal (least squares) estimate of the original source. This approach provides an approximate way to increase the effective order of the decoder, yet while retaining quite manageable complexity. The new approach is demonstrated to significantly improve upon standard MMSE-based JSC decoding performance, both for the case of nonpredictive source coding (e.g. vector quantization) as well as for predictive source coding (DPCM).