Algorithm Research On Parameter Estimation Of Non-stationary Sources

Slepian-Wolf coding theorem and Wyner-Ziv coding theorem lay for distributed sourcecoding systems the theoretical foundation, where an important hypothesis is always made. Inthis case, correlation information between correlated sources is supposed to be known by theencoder. However, in most circumstances, correlation information remains unknown or variesover time. Therefore, the above hypothesis goes against the real environments, which finallyleads to low coding efficiency. In order to solve this problem, this thesis gives deep researchon two different correlation estimation methods. One of the methods which can be used in theconventional source coding fields is based on hidden markov model, while the other onewhich can be exploited in the distributed source coding fields is based on sliding-windowbelief propagation algorithm. The main works are as follows:(1) Firstly, the non-stationary source is modeled as a hidden markov process. Afteranalyzing all parameters of hidden markov model itself, this thesis constructs a special hiddenmarkov model suitable for the thesis’s main purpose. Then the forward algorithm and thebackward algorithm are realized respectively to generate variables used for correlationestimation. After that, a mixed algorithm is created by jointly using results produced byforward algorithm and backward algorithm. Finally, the statistics of non-stationary sources isestimated. Experimental outcomes prove that the mixed algorithm is the best among the abovealgorithms in terms of estimation results.(2) Secondly, the encoding and decoding principle of low-density parity-check codes andlow-density parity-check accumulated codes are researched respectively. After that, the binarysymmetric channel is used to model the stationary correlation information. And four differentlow-density parity-check accumulated codes are utilized to encode sources while the standardbelief propagation algorithm is run to decode sources. In the end, the encoder manages tocompress non-stationary sources by using the estimation results generated by paramererestimator based on hidden markov model.(3) Finally, this thesis gives a detail explanation of how to use the sliding-window beliefpropagation algorithm to estimate the statistics of non-stationary sources and then to compressthe non-stationary sources. After that, two regular low-density parity-check accumulated codes are used respectively to test the sliding-window belief propagation algorithm in termsof estimation ablity and decoding performance, which is then compared with those producedby standard belief propagation algorithm.