Research On Virtual View Rendering Of Multiview 3DTV

Motivated by the increasing requirements of the immersive visual experience,the traditional 2D images are difficult to meet the growing requirements.3D technology is explored to reconstruct the 3D scene.In recent years,with the development of 3D technology,standardization and the popularity of 3D display terminals,3D technology is opening up a promising blue ocean market for entertainment,medical and education,etc.Naked-eye 3DTV and multi-view display lead users to a stereo world without wearing glasses,which are suitable for homes and advertisements.Researchers are paying more attention to it.3DTV system generally consists of 3D video acquisition,encoding,data compression and transmission,rendering,and 3D visualization.This thesis focuses on content rendering,which is the last step before displaying.High-quality virtual view image generation is a key technology to the depth-based 3DTV system.Besides,there is an inevitable problem in multi-view display,that is.people suffer from the abrupt view-image change on the border of neighboring viewing zones,which leads to dazzling discomfort.In this work,we propose a novel scheme based on DIBR and automatic foreground extraction.It improves the viewers'visual experience by disparity control and generates high-quality virtual viewpoint images.In summary,the main highlights of this thesis are:1.Foreground object extraction aids in disparity control and texture synthesis.The conventional RGBD image segmentation is combined with depth-learning-based object detection to automatically extract foreground objects.Faster-RCNN is employed to predict the bounding boxes of persons.With the boxes and the depth map,we select the seed point,which is input to Mean Shift algorithm to achieve accurate segmentation.This method improves time efficiency compared to manually selecting seed points.2.We propose a disparity control method to address the issue of the visual discomfort in multi-view display.We separately perform global disparity remapping and local depth adjustment for foreground persons.In consequence,the intermediate viewpoint gives a strong stereoscopic impression,which gradually weakens on the border of neighboring viewing zones,resulting in a smooth transition effect.3.We present an exemplar-based algorithm to handle the disocclusions.Leveraging on the segmentation results,we modify the priority tenn and find source patches within an expanded search window.Experimental results demonstrate that our method generates high-quality virtual view and surpasses the comparative methods in both subjective quality and external evaluation.Firstly,we introduce the background of virtual viewpoint generation,the development and the current research status of 3DTV at home and abroad.Then,in accordance with the flow chart,we elaborate the related work and the proposed methods.Finally,the research results are summarized and future prospect is put forward.