Filter Design And Coding Prediction Methods For 3D Video

3D video is a hot research topic in multimedia. In recent years, the depth map which reflects the distance between an object and the camera is introduced into conventional 3D video, and a depth-based 3D video system is established. With virtual view synthesis technique, a virtual video of any viewpoint can be rendered in such system. Through view synthesis, for each pixel, its disparity vector can be obtained with the depth information. Then, the pixel can be projected into the virtual view image. As the noisy depth pixel will cause the disparity error and lead to geometric distortion in the rendering result, noise reduction is a main research topic of 3D video. Besides, as the popularization of 3D video is restricted by its huge amount of data, many disparity-based interview prediction methods are proposed to improve the coding efficiency. However, these methods are still need to be optimized. Accordingly, some research works are presented in this thesis.(1) Depth spatial filter: First, this part indicates that the relationship between pixels is determined by the correlation of their corresponding points in 3D space. Accordingly, the pixel correlation is reformulated with three aspects, 3D-spatial distance, texture similarity and motion uniformity. Then, a 3D-spatial-texture-motion trilateral filter(3DSTM-TF) is proposed. Furthermore, for utilizing the chrominance information efficiently, texture images are converted from YUV color space to RGB color space. With the reference pixels which are selected based on the texture similarity in RGB color space, a depth boundary correction filter(BCF) is put forward.(2) Depth spatial-temporal filter: Based on the spatial filter, the relationship of pixels in the spatial-temporal domain is determined with 3D-spatial distance, texture similarity and time distance. Then, we proposed a 3D-spatial-texture-time depth spatial-temporal filter. As it is impossible to judge whether texture and depth is more reliable through filtering, a pixel vector is introduced with jointly utilization of texture and depth information. The pixel relationship is determined by the pixel vector similarity. Then, pixels which take part in the final filtering process are selected by considering the pixel vector similarity, and the denoising result is obtained with a robust median filter. Consequently, a pixel vector based depth spatial-temporal filter is developed.(3) Depth-based block partition(DBBP) signaling optimization: This part represents the DBBP mode in 3D-HEVC, and points out that unnecessary flags(conditions of never using DBBP mode) are written into the bit-stream. To solve this problem, a bug-fixing method of DBBP signaling is presented based on the partition size of DBBP.The abovementioned spatial and spatial-temporal filters have achieved good denoising performance, and enhanced the quality of synthesized results. Besides, the proposed bug-fixing method of DBBP signaling can improve the coding efficiency effectively.