Study Of The Correlation Noise Model Based On Distributed Video Coding

Traditional video compressing standards, e.g. MPEG or H.26x, employ a complexmotion estimation framework in their encoders, so they usually include ahigh-complexity encoder and a low-complexity decoder. This architecture restrictes itsusing in new applications emerging in recent years, such as sensor networks, videosurveillance systems, multi-view video and other fields. In order to meet thisrequirement, a new video coding scheme, distributed video coding (DVC), wasproposed because it has a low-complexity encoder and error resilient with highcompression efficiency. It has attracted a lot of attention and become a very hot researcharea.This thesis focuses on correlation noise model(CNM) of DVC aiming to designmore accurate CNM and provides effective information to reconstruction. The maincontributions of this thesis are as follows: First, an adaptive CNM generation algorithmis proposed. By considering the original information from encoder, the statisticalproperty of the transform domain correlation noise and the motion characteristic of theframe, the algorithm adaptively chooses Gaussian distribution or Laplacian distributionas CNM. For packet loss of channel, the components of correlation noise are analyzed,and CNM based on hybrid probability density function is proposed. Then,expectation-maximization(EM) algorithm is used to estimate the model parameters.Finally, a undirectional distributed video coding(UDVC) system based on transformdomain is constructed using the above-mentioned algorithms.Experimental results show that compared with H.263intra-frame coding, UDVCincrease PSNR4~6dB, compared with H.264intra-frame coding, UDVC increasesPSNR1~3dB. The reconstruction video quality is obviously enhanced by UDVC,besides, the proposed system whose encoder is simple and easy to realize, can beapplied to the video communication systems for low-complexity encoding terminals.