Research On Depth Coding Technology Based On View Rendering

Free Viewpoint Video/Televison(FVV/FTV) can provide the stereoscopic video with live immersion and allows the users to choose the watch position in a certain range freely, so FTV has become the next important research interest in the field of video signal processing. In order to achieve the desired applicaiton, FTV need extra depth map with traditional 2D texture video to generate virtual viewpoint. Compared with texture video, depth map has completely different properties and applicaitons. If the traditional video coding method is directly applied for depth map, it will cause large distortion for synthetic viewpoint. So how to effectively code depth map while preserving the quality of synthetic viewpoint has become a current research hot spot.Firstly, the thesis does a statistics for the utilization frequence of some key depth map coding technologies, where the statistics is based on 3D video codec international standard platform HTM. Then, two main research fields are determined, namely, Segment-wise Direct Component(DC) Coding(SDC) and Depth Modelling Modes(DMM). To improve the quality of synthetic viewpoint, we explore some effective coding methods for depth map in SDC and DMM respectively. The thesis contributes the following achievements:1. Set up an efficient optimization framework for SDC by two-stepIn the first step, for DLT-induced error in SDC process, we propose one single sample predition instead of four-corner prediction, and create an optimization model for SDC prediction and coding. In the second step, to improve the quality for synthetic viewpoint, we propose a view synthesis optimization method when searching an optimal SDC residual value at the encoder. Finally, we set up an efficient SDC optimization framework for prediction, coding and residual selection by two-step. Experimental results show the proposed SDC optimization model can bring coding gain up to 0.31%. For the standard study, one method will be proposed if it can bring 0.10% coding gain.2. Do optimizaiton for DMM1 and DMM4 among depth modelling modesFirstly, DMM1 contains a lot of wedgelet templates in lookup table, while templates stored in the buffer memory may lead to many storage problems such as cache misses. Through analyzing the generation process of lookup table for different size block, we propose the down-sampling method and establish a simplified model for wedgelet template lookup table based on up/down sampling. Experimental results show the proposed simplified model can achieve storage size reduction by 27.1% with negligible coding loss.Secondly, DMM1 is only applicable to depth block with straight line boundary, and DMM4 is mainly advantageous for cases that texture and depth are strongly correlated. For the depth block which contains non-linear boundary and is not correlated to texture, we propose the second-order wedgelet prediction to make up the shortage of DMM1 and DMM4. Experimental results show the proposed second-order wedgelet prediction can bring coding gain by 0.05%.