Research On Deep Image Coding And Filtering Technology In 3D Video

With the continuous promotion and development of 3D video technology, 3D film and 3DTV have been the essential consumer goods of people’s daily life.Recently, popular MVD format-based 3D video representation arouses wide attentions, which includes multiview color videos and depth videos with its camera parameters. Owing to its additional depth videos except for traditional multiview color videos, this format imposes a heavy burden on storage and transmission. Although video coding techniques can reduce the amount of data,traditional color video coding could not satisfy depth video coding, especially with great quantization parameters, which produces severe distortions of depth image along the boundary regions. This will directly result in wrong 3D warping and finally badly destroy the 3D visual effects. In order to reduce the boundary distortion of coded depth image resulted from traditional coding, this thesis mainly focused on the research of depth video coding methods and depth filtering for compressed depth images. The main works and innovations are as follows:1. a method of depth image scalable coding based on edge’s lossless compression is proposed. In the proposed scheme, a threshold is firstly deduced based on synthesis-guided for edge detection and edge detection is conducted based on this threshold. Then, scalable stream structure based on edge and down-sampled image is proposed. At last, the depth map’s reconstruction is achieved based on partial differential equation method. Experimental results show that proposed scalable coding scheme has better rate-distortion performance than other’s method and the synthesized virtual view’s quality has been improved greatly. Meanwhile, the boundary regions of depth image and virtual image can be well protected.2. candidate value-based depth boundary filtering scheme is proposed. The reliable pixels and unreliable pixels are firstly defined and propose that only detected unreliable pixels along the boundaries are filtered, which greatly reduce time of filtering. Then, depth image’s spatial correlation and statisticcharacteristic can be used to achieve the filtering and enhancement of distorted boundary regions for depth image, so proposed scheme utilizes these features to determinate the candidate values. Finally, boundary filtering of depth image is achieved by minimizing the absolute difference between current pixel value and its candidate ones in the candidate pixels set. Experimental results show that the proposed scheme has better synthesized quality for virtual view than ADTF 、JTF、DBRF and our visual qualities are better than others.