Joint optimization of source and channel coding based on a nonlinear estimate receiver

This thesis considers the problem of joint optiniization of source and channel coding algorithms for communicatio~is over the additive white (;aussian noise (AWGN) and tlie Rayleigh fading channels. In the proposed system, the analog source signal is first co~npressed by a vector quantizer (VQ). The output of the VQ (tlie VQ index) is mapped directly into a signal vector in the modulation signal space, and the signal vector is transmitted over a noisy channel. A receiver based on a nonlinear conditional estimate is used to reconstruct a replica of the source signal directly from the received signal. The main blocks to be optimized in the joint source and channel coder are the VQ encoder, tlie napping fro111 the VQ index to the modulation constellation, the ~nodiilation sonstellation, and tlie decoder structure. Subject to constraints on the average energy and bandwidth, the objective is to mini~nize the mean-square error (?VISE) between the original and tlie reconstructed signal. Based on Bayes estimation theory, a soft decision vector quantizer (SDVQ) was developed. The optimal decoder for this system computes the conditional ~riean of the source signal given the received channel signal. The output of such an optimal decoder is a linear combination of the VQ centroids for the SDVQ partition, in whish the weighting coefficients are nonlinear functions of the received signal. Several approximate i ~ ~ l p l e ~ ~ l e ~ l t a t i o ~ i s t various channel SNR were also studied. An iterative algorithm is presented for the joint design of the VQ and the modulation signal set. The algorithnl first opti~nizes tlie VQ codebook for a fixed signal set, and then optimizes the signal set for a fixed VQ codebook. Iterating these two steps imtil convergence occurs will provide at least a locally optimal solution to tlie probleni. The algorithnl was used to design tlie VQ and signal constellation for a first order Gauss-Markov source operating in the AWGN and Rayleigh fading channels. The simulation results indicate that the system performance is significantly enhanced by the joint design, especially when the channel signal-to-noise ratio is low. The i ~ n provement in the signal-to-noise ratio (SNR) for the reconstruc-ted signal can be up to 5 dB. Due to the constraints on the VQ encoder delay arid VQ complexity, the source coder c-an not remove all the redundancy in the source. The residual redundant-y is nodel led as a first order Markov process. We further developed a sequential decoding algorithm to exploit the residual redundancy in order to itliprove the performance in noisy channels without any bandwidth expansion. The simulation results show that significant improvement can be obtained by using the sequential decoding strategy, especially in a Rayleigh fading channel.