Multiple Description Coding For 3D Depth Image

Three dimensional stereo video is a new type video technology that can provide a three-dimensional sense. It is popular with people because of its strong presence sensation. At present, multi-view video plus depth video has become one of the most popular forms of 3D video expression. Depth map compression has been an important part of the new generation of 3D video coding.3D depth image has its unique characteristics, the traditional 2D image video coding method does not apply to the depth image, it is necessary to seek depth image coding scheme. In addition, because the inherent visual features of HVS can ignore a lot of perception redundancy, so if the visual characteristics of the HVS can be used in the image coding, we will get better visual effect.In this thesis, based on the traditional multiple description lattice vector quantization coding scheme, the coding algorithm is studied and improved according to the characteristics of the depth image and the visual characteristics of HVS. The main work includes:(1) An optimized coding scheme based on multiple description lattice vector quantization(MDLVQ) for depth image is proposed. Given the sparse structure characteristics of depth image, MDLVQ is applied in an optimized way. In the scheme, the proportion of the edge for each image block is used as the principle to the adaptive allocation of quantization step size. For the block with more edge information, a smaller quantization step size is used, so that we can realize more accurate coding for edges blocks, improving the quality of the reconstructed image and the synthetic virtual perspective eventually.(2) A new JND model is proposed, which includes depth information. Traditional single view video coding has taken spatial JND model into account, but the human visual system can also notice the change of depth information. The scheme combines the SJND and DJND, getting a new JND model namely SDJND model. The new JND model considers not only the hidden effect in space domain, but also considers the visual redundancy in depth image. Compared with the SJND without considering the depth information, the SDJND model can better match the visual perception redundancy.(3) A multiple description depth image coding based on the human visual system is proposed. The DJND model extracted from SDJND is applied to 3D depth image coding. We allocate the quantization step size according to the percentage of the difference between the reconstructed depth image and the original depth image. The proposed model can get better visual quality in the case of the same bit rate.