Soft trellis waveform compression and joint source/channel coding

This thesis aims at improving the performance of trellis waveform compression systems and joint source/channel coders. We first provide the theoretical framework for the new concept of soft source coding, which is based on exploiting the similarity between trellis source compression and maximum likelihood channel decoding. The new approach involves using the maximum a posteriori soft-output channel decoding algorithm for trellis source compression. Given a block of source output vectors, the new algorithm delivers a set of soft reliability values that describe the likelihood of the different symbols at the encoder output, in the minimum distortion sense. As a trellis search algorithm, the new soft procedure performs as well as the Viterbi algorithm.; The soft reliability information delivered during compression is employed to derive a fuzzy trellis codebook search algorithm. The new algorithm outperforms the Linde-Buzo-Gray (LBG) algorithm in delivering lower distortion reproduction codebooks. Unlike the LBG algorithm, the new fuzzy approach is significantly less sensitive to the initialization phase, and provides improved performance using “short” training sequences. This feature is particularly of interest with practical nonstationary sources of information. The success of this fuzzy algorithm is extended to the case of channel-optimized trellis waveform compression. Over a practical range of channel bit error rates, and compared to the channel-optimized LBG algorithm, the new approach provides lower distortion channel-optimized codebooks irrespective of initialization using short training sequences. Furthermore, the optimal channel-optimized codebooks show improved performance under channel-mismatch conditions.; The last part of this thesis addresses the design of joint source/channel (JSC) soft decision decoding/detection systems. First, we study the performance of soft decision JSC decoding using k-DPCM compressed images over memoryless Gaussian channels with convolutional channel codes. Then, we show the advantage of using higher redundancy levels in providing higher coding gains, especially during “bad” channel conditions. This advantage is highly utilized when the alphabets of the source encoder output match the channel encoder input. For this purpose we test the performance using dual-k convolutional codes and k-DPCM systems. Furthermore, we provide a comparative study between sequence-MAP and symbol-MAP JSC detection. Although the two approaches provide similar bit error rates, we show that symbol-MAP detection is always associated with lower average distortions of reproduction as applied with VQ systems.