Research On Stereoscopic Video Coding Based On Visual Saliency

Stereoscopic video is a collection of multiple video sequences, which capturingthe same scene via two-view or multi-view camera. Compared with the single-channelvideo, stereoscopic video increases the depth information of the scene, and enhancesthe three-dimensional sense and realism of the video. Stereoscopic video has notwidely used in practical applications yet since it has a large amount of data. Therefore,the efficient compression of the video data is significant for reducing the need forstorage and transportation. In order to achieve a certain degree of compression, theprediction and compensation methods for intra-frame, inter-frame and inter-view areemployed to exploit the temporal redundancy, spatial redundancy and inter-viewredundancy, which are abundantly existed in stereoscopic video. However, theconventional stereoscopic video coding approaches ignore the diversity that humanvision system (HVS) perceives on video scene. This thesis starts from the visualcharacteristics of the human eyes and studies stereoscopic video coding methodsbased on the visual saliency.A large number of physiological and psychological researches based on HVShave shown that while watching and comprehending video images, human will focusattention on several regions which may cause visual saliency. These regions canexcellently represent the scene information and image content, they are worthconcerning and are called region of interest (ROI). According to the differentinteresting degree and different sensitivity to distortion of different regions, depthinformation is adopted in this article to divide the video sequences to multi-level ROI,and appropriate measures are taken to adjust quantization parameter (QP) so that morebits are allocated for ROI to improve the video quality, and fewer bits for non-ROI toachieve more efficient compression. Multilevel region of interest guided bit allocationmethod can enhance the smoothness of ROI boundary, degrade bit rate while ensuringthe overall quality of the reconstructed video, and improve the efficiency ofmulti-view video coding.Multi-view video allows users to freely change their viewpoints and enhances3Dvisual experience, but it is difficult to transmit multiple viewpoints in resource limitedenvironment. Texture and depth data can be utilized to generate the multi-view video required by stereoscopic display device in virtual view rendering technology. Thisdissertation studies virtual view rendering oriented multi-view video plus depthcoding (MVD) method based on the theories of visual saliency and virtual viewrendering. A ROI extraction method based on depth, motion and depth edge as well asan optimized coding method is proposed for texture video and depth video.Experimental results show that the optimized method can enhance the quality of ROIin the virtual viewpoint without increasing the coding bits.