Research On Rate Distortion Optimization Of Depth Coding For Three Dimension Video

Three dimension video(3DV) has attracted extensively attention recently since it can provide viewer with the vivid visual perception about the scene. To meet the demand of broad 3DV applications, the Moving Picture Experts Group(MPEG) launched the 3D video coding project in 2011, covering solutions based on high efficient video coding(HEVC). Multi-view Video plus Depth(MVD) format has been chosen to be the recommended data format of the HEVC based 3D video coding standard(3D-HEVC) since it has the advantages of good compatibility and excellent rendering flexibility. The additional depth map of MVD can synthesize the virtual view freely by using depth image-based rendering(DIBR) technique.The rate distortion optimization of video encoding system is based on the theory of rate distortion optimization to select the optimal parameters of encoding. Because the reconstructed depth map is used to synthesize the virtual view instead of the direct viewing, 3D-HEVC employs the rate distortion optimization based on virtual view distortion to select the best encoding parameters of depth map. View synthesis(VS) based distortion models, synthesized view distortion change(SVDC), provides high compression efficiency while imposes a heavy computation burden. In order to reduce the complexity of depth map encoding, two solutions are investigated in this thesis.A new model based virtual view distortion prediction algorithm is firstly proposed in this thesis. At the beginning of coding each depth map, VS is executed with the original texture map and the original depth map to record the category of each pixel. Pixels are classified into four models: non first occlusion model, first occlusion model, interpolation model and disocclusion model. Then, during encoding each depth block, the distortion of each pixel is calculated by different model by the corresponding category. Finally, the distortion of each pixel is added as the distortion of the current virtual view. We integrate the proposed algorithm into the 3D-HTM reference software 11.0. Compare with other synthesized view distortion estimation methods, the proposed algorithm has the lightly change of coding time and better coding performance.A new algorithm of progressive rate distortion optimization is also proposed, which is based on the synthesized based view distortion optimization. We define the three early termination conditions into three cases: bit aware termination, depth-distortion aware distortion and rendering-distortion aware termination. The current RD cost is updated and compared with the existing minimum RD cost in the three cases. If the current RD cost is larger than or equal to the existing minimum RD cost, we should stop calculating the RD cost of the current test block and eliminating the current mode. Remarkably, during the rendering-distortion aware termination, the current RD cost is updated and compared after each rendering process of multiple virtual synthesis position. The proposed algorithm brings 0.1% performance loss while 36.9% depth coding complexity can be saved compared with HTM11.0.