On the design of robust vector quantizers for sources and channels with memory
Miroslaw Pawlak, Pradeepa Yahampath · 2002
Recently, robust quantization has received considerable attention, particularly as a potential approach to joint source-channel coding. This dissertation investigates the practical design of channel optimized vector quantizers (COVQ). Specific problems considered here include COVQs with memory and COVQs operating over channels with memory . In this work, the emphasis is placed on soft-decoding at the receiver. Vector quantizers with memory are an effective means of quantizing correlated signals. However, when designed without regard to channel errors, these quantizers suffer from degradation of performance due to the propagation of channel errors at the receiver. We consider two important examples of such quantizers, namely, predictive vector quantizers (PVQ) and finite-state vector quantizers (FSVQ). In the case of PVQ, an iterative algorithm is developed for jointly optimizing the quantizer and the associated linear predictor to a given channel. According to the simulation results presented here, the proposed PVQ designs based on hard-decoding perform comparably to those obtained by a previously studied gradient-search optimization algorithm. Furthermore, it is demonstrated that PVQs with soft-decoding can provide a significant improvement over hard-decoding systems. In the case of FSVQ, a time-recursive decoding algorithm, which exhibits graceful degradation of performance with increasing channel noise, is introduced. Design of channel optimized FSVQ is also considered. Simulation results are presented, which demonstrate that proposed channel optimized FSVQs outperform the memoryless COVQs operating at the same rate. Finally, in the context of channels with memory, joint equalization and soft-decoding using a sliding-block decoder is investigated. This decoder is a non-linear time-invariant filter based on minimum mean square error criterion. As a practical implementation, multi-layer perceptron (MLP) is considered. Simulation results indicate that MLP-based soft-decoder outperforms a previously studied recursive soft-decoder, particularly under high channel noise. However, the complexity of the optimal sliding-block decoder function is found to increase with the encoder resolution, making the estimation task harder. In an encouraging development, it is shown that the optimal sliding-block decoder for the Gaussian channel approximates a linear function as the channel becomes noisier. Experimental results seem to support this theoretical result.