Researches On Content-based Multi-View Video Coding

Multi-View Video (MVV) with features such as interactive, editable and stereoscopic perceptual which can provide a more vivid visual experience becomes more and more popular, it has attracted more and more attention from the industrial sector and the academic, gradually become the next-generation multimedia systems'key technology. But the multiplied data challenges its storage and network transmission, which greatly limits the practical applications for Multi-View Video. Therefore the Multi-View Video Coding (MVC) technology aimed to reduce the bits for storage and network transmission has been a key and focus in this field. This dissertation researches MVC based on content, it pulls essential factors such as motion objects, regions of interest in coding to improve the coding efficiency. The main results achieved include the following sections:(1)The depth information of MVV is conducive to video object detection because it is very close to the semantic information. This dissertation proposes a method for motion detection based this feature. The method converts a sequence of consecutive depth images into a sequence of consecutive temporal-to-spatial slices in the horizontal and vertical directions. In each one temporal-to-spatial slice, semantic information and motion information are both included, the background region forms a vertical line pattern, and a moving object creates an irregular, non-vertical structure. Then binarizing the temporal-to-spatial slices by a dynamic threshold,reconstructing the converted slices in the horizontal and vertical directions into temporal image masks and reserving the common parts of two corresponding masks,finally post-processing the reconstructed images. Experiment results show that the proposed method exhibits a good performance for motion detection.(2) This dissertation proposes a fast MVC algorithm based on motion detection that divides the MVV images into moving regions and non-moving regions. In these two kinds of regions, there are different distributions for the macroblock encoding modes and motion predication methods. The moving regions tend to choose more time-consuming parts of the coding strategy as the optimal coding strategy, and the non-moving regions are on the contrary. This dissertation improves the coding strategy from the point that the optimal macroblock encoding modes and the optimal motion predication methods are different between the moving regions and non-moving regions in B frames. This coding algorithm can achieve about 67% reduction of encoding time in comparison with JMVM7.0, while it hardly influences the rate distortion performance of MVC. (3) This dissertation proposes an algorithm for MVV's bits allocation based on Region-of -Interest (ROI). A large number of human vision physiology and psychology experiment results show that the human eyes on the different aspects of the video image are not equal, but to show the selectivity and migration, the visual masking makes some changes occurred in the regions in which human eyes are not interested imperceptible. Based on this theory, this dissertation distinguishes ROIs using color, intensity, motion, orientation and depth information according to the Bottom-to-Up visual model. Vision Priority Value (VPV) is used to represent the ROIs, if the value is greater and the region is more attractive, so more bits are allocated in this region. Experiment results show that the proposed method extensively improves data compression efficiency. Comparing with JMVM7.0, it can achieve 18%~34% bits saving while the subjective image quality remains almost intact.