Analog joint source-channel coding for non-standard scenarios
Date
2015
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Recently, analog joint source-channel coding (JSCC) systems based on mappings have become one of the most promising schemes for transmitting discrete-time, continuous-amplitude sources (e.g., audio and video samples) over time-varying channels (e.g., wireless channels) under complexity and delay constraints. In contrast to traditional digital communication systems based on Shannon's separation principle, analog JSCC schemes are robust to changes in the channel quality, and do not require near infinite block lengths to approach the theoretical limits. As a result, the encoding/decoding complexity and delay can be greatly reduced compared with digital schemes. Direct source-channel mappings take K discrete-time, continuous-amplitude symbols (a K dimensional vector in the source space) and map them directly into L discrete-time, continuous-amplitude channel symbols (an L dimensional vector in the channel space), achieving either bandwidth reduction (K > L ) or bandwidth expansion (K Direct source-channel mappings take K discrete-time, continuous-amplitude symbols (a K dimensional vector in the source space) and map them directly into L discrete-time, continuous-amplitude channel symbols (an L dimensional vector in the channel space), achieving either bandwidth reduction (K > L) or bandwidth expansion (K < L ). Most of the work on these schemes has dealt with the transmission of memoryless sources over noisy channels for point-to-point communications. In this dissertation, we focus on distributed scenarios and non-i.i.d. sources. Specifically, we first study the problem of transmitting multivariate correlated Gaussian samples over AWGN channels. A direct source-channel mapping designed utilizing power constrained channel optimized vector quantization (PCCOVQ) is proposed, taking into account the channel noise and power constraints. Simulation results show that, for bandwidth reduction, direct source-channel mappings achieve the theoretical limits for low channel signal to noise ratios (CSNR) when the samples are correlated. The performance is also quite close to the theoretical bound for higher CSNR. Second, we work on the design of direct source-channel mappings using space-filling curves for the transmission of memoryless Gaussian samples over AWGN channels when side information is available at the receiver (Wyner-Ziv scenario). We first propose a 1:1 mapping for K = 1 and L = 1 using a periodic piece-wise linear curve. Then, we propose a 1:M bandwidth expansion mapping based on the use of existing space-filling curves in a periodic fashion. A simplified decoding algorithm is also proposed to reduce the cost of decoding without losing much performance compared to MMSE decoding. Then, we propose a flexible rate K : L bandwidth expansion mapping for the Wyner-Ziv scenario by combining the proposed 1 : 1 and 1 : M schemes with the optimum energy allocation. These schemes are shown to outperform existing systems for a wide range of CSNRs, especially for high CSNRs and highly correlated side information. We also study analog JSCC for the transmission of correlated Gaussian senders transmitted over separated AWGN channels. An asymmetric distributed coding scheme is proposed. One of the senders is encoded using standard mappings and the reconstructed symbol at the common receiver is used as distorted "side information" for the other sender, which is encoded by a periodic mapping designed for the point-to-point Wyner-Ziv scenario. A power allocation strategy is proposed to minimize the distortion while still satisfying the power constraint. Simulation results shows that the proposed schemes perform very close to the theoretical limits, outperform existing analog zero-delay mapping schemes, and are robust against CSNR mismatch and correlation mismatch. Finally, we extend our research on analog JSCC to multiple access channel (MAC) scenarios. The proposed scheme contains a CDMA-like access scheme which converts the MAC into orthogonal channels. We showed that the proposed CDMA-like access scheme achieves capacity if optimum power allocation is used. Simulations show that by pairing with periodic mappings, the proposed scheme performs very close to the theoretical limits irrespective of the rate of each user when independent Gaussian sources are transmitted over a Gaussian MAC with side information at the receiver. When transmitting correlated Gaussian sources over a Gaussian MAC, the performance of the proposed scheme is comparable to that of the best existing zero-delay schemes, but with the benefits of easiness of adaption to different rates.