| Summary: | This thesis proposes an optimally rate allocated image transmission system that uses
Vector Quantization (VQ) for source coding, and a family of variable rate Punctured Convolutional
Codes (PCCs) for channel coding. At the receiver, we apply the source aided channel
decoding technique known as Markov Model Aided Decoding (MMAD). Our optimality criterion
is to maximize the average end-to-end Reconstruction Signal-to-Noise Ratio (RSNR) under the
constraints of a fixed information rate (in pixels per second) and a fixed transmission bandwidth.
For a given channel SNR, this joint source-channel coder design achieves the optimal rate allocation
between the source coding and the channel coding operations. Compared with the conventional
rate allocated system (the analogous system that does not use MMAD), the proposed
system gives significant performance improvement. This is due to the fact that MMAD increases
the strength of the channel codes, thus allowing the system to allocate more rate to the source
coder, which results in a higher resolution image.
In the course of our study, we first investigate MMAD without explicit channel coding for
VQ image transmission over the noisy memoryless channels comprising the Binary Symmetric
Channel (BSC) and the Additive White Gaussian Noise (AWGN) channel. In order to evaluate the
effects of the order of the Markov model of the data, we consider two types of decoding
algorithms. One is based on the Viterbi sequence decoding algorithm, the other is based on the
Bahl, Cocke, Jelinek and Raviv (BCJR) decoding algorithm. The former is computationally less
complex, and is optimal (in the sense of minimizing the Bit Error Rate (BER)) for decoding with
a first order (O(l)) model; while the latter allows an efficient, but slightly sub-optimal, decoding
algorithm for decoding with a second order (0(2)) model. We find that most of the MMAD
coding gain is already achieved by using the 0(1) model, and therefore in the remainder of the
study consider it only.
We go on to analyze two types of O(l) MMAD with Convolutional Codes (CCs)
employed for explicit channel coding. We call the decoders Markov Model Aided Convolutional
Decoders (MMACDs), and show via simulation that the performance benefits attained by using
the Markov model are similar to the large gains found for MMAD without explicit channel
coding. One type of MMACD is based on the Viterbi algorithm, and applies a trellis merging
technique. This decoder has an optimal BER performance, but has the constraint that the length of
the source codewords be less than the memory of the CC. The other MMACD is a concatenation
of a soft-output channel decoder followed by an MMAD without channel coding. This decoder
does not have the constraint on the length of the source codewords, but has less coding gain than
the trellis merged decoder.
Finally, we investigate the problem of optimal rate allocation between the source coding
and the channel coding for VQ/PCC transmission systems that employ MMAD. Our simulation
results over the AWGN channel show that the optimal rate allocated system is superior in RSNR
performance to the optimally rate allocated system without MMAD. The MMAD coding gain
depends on the image, but is typically 2 dB in channel SNR. We find that for the conventional
system, the point of optimal rate allocation is fairly independent of the image; while for the
MMAD system the allocation depends strongly on the image characteristics. Because of this, the
rate allocation calculation is significantly more complicated when using MMAD. The rate
allocated systems require an estimate of the channel SNR. Because in practice there will always
be some inaccuracy in estimating this, to conclude our study we investigate the sensitivity of the
rate allocated systems to channel mismatch, and find them to be fairly robust. === Applied Science, Faculty of === Electrical and Computer Engineering, Department of === Graduate
|
|---|
