Multi-view Video Coding Based On Correlation

More and more people who work on science research devote themselves to the research of multi-view video code with the 2D video changing more and more maturely. By analyzing the characteristic of the multi-view video data, we designed a correlation testing structure. And two test methods are also proposed, which are the correlation model based on block matching and the correlation model based on the correlation functions. According to the correlation intensity on both the temporal and spatial direction, a method of multiple prediction structures coding jointly is proposed. It analyze the correlation of the sequences at first and judge which kind of correlation feature the sequence is. And based on the results, three flexible prediction structures are designed by utlizing the reference pictures between views pertinently. Experimental results show that combining three prediction structures together can reduce a lot of coding bitrate by sacrificing only a little PSNR. And to the sequences whose spatial correlation is quite strong, the prediction structure given can not only raise the PSNR, but reduce the bitrate as well.Some researches have also been made on the prediction structures whose spatial correlation is strong and a revised prediction structure is proposed. As the most important feature of the sequences whose spatial correlation is strong is that the distance between neighboring cameras is short, so the disparity between neighboring cameras is also very small. Even if there are several cameras between them, the global disparity increases little. So we can improve the coding efficiency by adding the number of B-views. The experimental results show that the PSNR of the revised prediction structure can improve 1.3 to 1.5dB compared with the MVC-HBP structure under the same bit rate. And it can also save about 80 to 120kbps compared with the MVC-HBP structure under the same PSNR.In addition, the random access of multiple prediction structures coding jointly is also assessed. The experimental results show that the random access of multiple prediction structures coding jointly is superior to that of MVC-HBP. What's more, the position of the I-view of the three structures is set to the mid view, which can be referenced by two views. Compared with MVC-HBP structure, the other seven views can decode less pictures when it is accessed. The two points above improve the random access of the prediction structures on the whole.