Algorithmic approaches to joint source-channel coding.

摘要

A typical communication system includes two subsystems: source coding and channel coding. The goal of source coding is to remove redundancy from the source to utilize the communication channel efficiently and to reduce storage requirements; the goal of channel coding is to protect the source from channel noise by introducing controllable redundancy. By Shannon's source-channel separation theorem, the two subsystems can be optimized independently and performed sequentially without any sacrifice of optimality. The theorem, however, was developed asymptotically (using arbitrarily large coding blocks), and assuming that the channel condition is known and the communication is point to point. These conditions and assumptions seldom hold in practice. Practical systems of better performance can be built by the approach of joint source-channel coding (JSCC), in which the two subsystems are designed together rather than independently in tandem, and optimized simultaneously based on both source and channel characteristics.; The first JSCC technique to be studied is multiple description coding for robust transmissions over packet erasure channels. The basic idea is to create multiple descriptions of an original message, and deliver the descriptions independently through different routings. The receiver can reconstruct the message by any subset of those descriptions, and the reconstruction quality improves as the number of received packets increases. Reed-Solomon (RS) codes are used to correct channel erasure errors. We add uneven error protection (UEP) to consecutive segments of scalable source sequence with the redundancy strength of RS codes proportional to the importance of different segments. We study the problem of optimal allocation of RS code to protect scalable source sequence over packet erasure channels in the sense the expected reconstruction distortion is minimized.; In Chapter 3, we consider the maximum a posteriori (MAP) decoding of variable length codes over noisy channels. MAP detection and estimation is a useful tool in joint source-channel coding (JSCC), which exploits the residual redundancy remaining in the source code to correct/alleviate transmission errors even in the absence of channel code. We study the MAP decoding of variable length encoded Markov sequences over a binary symmetric channel (BSC) with or without the knowledge of the count of transmitted source symbols. Later, the noisy channel model is extended to a BSC with insertion and deletion errors, and a MAP decoding algorithm is proposed for such a channel.; In Chapter 4, we study the joint source channel decoding (JSCD) of VQ-coded two-dimensional signals like images. The basic idea is to scatter adjacent image VQ index bits into different packages, the packages are transmitted over packet erasure channels individually. At the decoder end, damaged VQ indexes are recovered by exploiting residual redundancy remaining in image VQ indexes. The straightforward MAP decoding in the two dimensional case has high complexity. To circumvent this we propose a MAP estimator exploiting residual redundancy in the two-dimensional case through high order context modeling that does not suffer from the problems of high time and space complexities and context dilution.; Finally Chapter 5 concludes the dissertation by summarizing main contributions and suggesting some interesting future work.