Correlation Estimation For Distributed Wireless Video Communication

Distributed source coding (DSC) studies separate compression and jointdecompression of multiple correlated sources. Its theory is based on Slepian-Wolftheory and Wyner-Ziv theory. Distributed Video Coding (DVC) is one of the earlistand most advanced applications of DSC to date. DVC utilizes the Wyner-Ziv theoryto avoid a computational expensive motion search at the encoder by turn theoperation to the decoder. In WZ coding, the encoder does not need to know or useside information, thus making it possible to accomplish predictive coding withoutencoder motion compensation. In WZ coding, the statistics of the correlation noiseare known to both the encoder and the decoder and does not change over time. InDVC, however, the decoder needs to generate side information using only theinformation it has. Regardless of how side information is generated, correlationnoise statistics will be unknown and dynamically change over time. In a nutshell,estimation correlation statistics has been identified as a key challenge in DVC.Usually, correlation error is modeled as a Gaussian or a Laplcian random variablewhose parameters are estimated from previously decoded frames. In the state-of-artwireless video communication technology DCAST, the encoder encodes the signalwith the help of DVC. The difference between DCAST and conventional DVC isthat in the former the encoder correlation noise is estimated by using the encoderside information and current encoding frame. To get the encoder side information,DCAST performs motion estimation (ME) and motion compensation (MC) at theencoder. This operation increases the encoding efficiency, and meanwhile increasesthe coding complexity significantly. DCAST also performs ME and MC at thedecoder to get the decoder side information, which is used with the decodedresidual together to recover current frame at the decoder. We can take the traditionalDVC’s coding mode, remove the ME and MC from DCAST, estimate thecorrelation noise by estimating the variance of current and its decoder sideinformation. On the other hand, the lack of the motion vector at the encoder makesit difficult to estimate the correlation noise. In this paper, we propose a correlationnoise estimation model to estimate the correlation noise without ME and MC at theencoder, and using the estimated noise to calculate the quantization step size in DCAST’s encoder coset coding. The main content and novelties are listed below：1. Proposed a model to estimate the correlation noise between source and decoderside information at the encoderWe propose a linear model to estimate the variance of correlation noise betweensource and decoder side information by referring the zero motion prediction at theencoder based on a Markov field assumption. More than that, in further experiment,we also take the noise of wireless transmission channel into account, and add thecomponent of channel noise to our model to make it more accurate.2. Proposed a new method to calculate the quantization step size in coset codingIn our new method to calculate the step size, we introduce the correlation noisewhich is estimated above, and determine the new quantization step size by analysisthe confidence interval of the residual.