Noise Model Research Based On DCT Domain Residual Distribution Characteristics In Distributed Video Coding

There have been a lot of achievements in the research area of Distributed Video Coding(DVC), for more than ten years developing. Compared with traditional joint-codingjoint-decoding video compression standards, Distributed Video Coding encodes each frameindependently at the encoder, and decodes them jointly using the correlation among the videoframes and channel decoding technology at the decoder, which can greatly reduce thecomplexity of encoding and improve the compression performance of the whole system. Thus,DVC is especially suitable for emerging applications, such as wireless video monitoring,multimedia sensor network, mobile videophone and camcorders, etc.Research on correlation noise modeling between original Wyner-Ziv(WZ) frame and sideinformation is one of the key technology in DVC, and it has a great effect on the compressionperformance of the whole system. Based on the basic principle and system framework ofDVC, this paper makes in-depth study on DCT subbands residual distribution characteristics,the main work and achievements are as follows:1. By deeply research, we found that the uncertainty of the noise distribution, and theattenuation characteristic of the DCT residual sub-band, result in that it is not accurate usingthe only one model parameter to describe the whole sub-band residual. Besides, using thevariance of residual coefficients to obtain the model distribution cann’t match the distributionfeatures accurately. Based on these characteristics, this paper presents a novel noise modelestimation method based on the simple classification of DCT sub-band residual coefficients.First, the residual coefficients in each sub-band are sorted according their absolute values;Then the most suitable correlation noise parameter estimation methods is chosen for eachresidual coefficient category. Simulation results show that the noise model parametersobtained by simple classification algorithm can match the residual coefficient distributioncharacteristics much accurately, and significantly improve the rate-distortion performance ofthe whole system.2. In this paper, we also find that the residual coefficient histogram is not entirely monotonous, but there exist a lot of small mutations, named small peak and trough, whichcoefficients. On the other hand, residual sub-bands in video frame are not isolated from eachother, they are correlated. In order to dig the related features between different residualsubbands, this paper put forward two correlation noise model estimation methods based onimproved FCM(Fuzzy C-Means), which are called the adjacent subband clustering and groupsubband clustering. In the two methods, the current subband to be decoded is clustered basedon the feature vector, and the corresponding parameter estimation method is chosen for eachclustered residual category. After the subband is decoded and reconstructed, the feature vectorof next decoding DCT sub-band is updated. The process is repeated until all the subbands arecompletely decoded. Simulation results show that compared with subband-level Laplaceestimation and presented simple classification method, this algorithm increases the decodingcomplexity, but it can describe the correlation noise statistical distribution characteristicsmuch more precisely, and can greatly improve the compression performance.