Research On Side Information Generation In Distributed Video Coding System

With the development of wireless network and people’s requirement for information, the multimedia communication services under mobile environment becomes the key business in the future wireless broadband network, such as wireless digital video camera, wireless sensor network, etc. Limited by computation capability, bandwidth, power consumption, real-time coding and so on, all these mobile terminal equipments need video encoder with very low complexity. Consequently, the traditional video coding standards such as MPEG and H.264/AVC are not suitable for these emerging applications. Distributed Video Coding(DVC) is a new video coding paradigm in which the computational burden is shifted from the encoder to the decoder. DVC has the feature of simple encoder and better error resilience, and provides an excellent solution for the above-mentioned applications.In DVC, the quality of Side Information(SI) strongly influences the Rate Distortion(RD) performance of the whole system. The more accurate SI is generated, the less parity bites are needed to be sent to the decoder, thus higher compression efficiency can be obtained. SI generation is one of the key technologies in DVC. This thesis first analyses the basic theory, related technologies and system framework of DVC systematically, and then focuses on the side information generation algorithm. The main contents are as follows:(1) Motion compensation based side information generation algorithm is researched, and an improved Motion Compensated Interpolation(MCI) algorithm based on variable size block motion estimation is proposed. Firstly, a coarse motion vector filed is obtained by forward motion estimation with weighted SAD criterion, bi-directional motion estimation in a small search range and a weighted vector median filter. Subsequently, the obtained motion vector is refined through using variable size block motion estimation from coarse to fine in a muti-stage strategy. And then the refined motion vector will be closer to the true motion trajectory gradually. Finally, high quality SI is generated through an adaptive weighted motion compensation technology. Experimental results show that the proposed algorithm has a higher performance than the existing algorithm.(2) Unsupervised forward motion vector learning method based on Expectation Maximization (EM) algorithm is one of the best SI algorithm in terms of RD performance, but it has the disadvantage of high computational complexity. To decrease the complexity, two optimization strategies are proposed:Motion prediction mechanism and Bilinear vector interpolation technology. In Motion prediction mechanism, the initial motion vector of every block in current Wyner-Ziv frame is predicted firstly, then the search range and the initial probability distribution model is adjusted adaptively. This strategy can speed up the E-step iterative learning process. In Bilinear vector interpolation, the resolution of the statistical motion field distribution is increased from blockwise to pixel precision, thus a more accurate soft SI which will further improve the M-step process could be obtained. Experimental results shows that compared with the original unsupervised motion vector learning algorithm with similar RD performance, the learning time of motion vector is reduced by means of the proposed method.