Research On Deep Image Coding Of 3D Video

3D video can provide the audience a more vivid and real perception, thus it attracts wide concerns and researches, In order to achieve their final impressions, a new 3D data format Multiview Video plus Depth(MVD) has been proposed, which including captured multiview video sequences and corresponding depth maps. This new video format can synthesize arbitrary virtual views by using depth image based rendering(DIBR). Hence, how to efficiently compress MVD data is an important issue in study of 3D video. Multiview video can be effectively compressed with traditional Multiview Video Coding(MVC) standard, but to depth maps, traditional method can make edge distortion. In this thesis, the research focuses on the compression and coding of the depth sequences in MVD, the main works and novelties are listed below:1. A depth map coding scheme used for arbitrarily shaped region is proposed. In this scheme, considering the characteristics of the depth map, a depth map is first divided into some arbitrarily shaped regions along the detected edges, and then the edges are coded without loss using a directional 8-connected chain code and an arithmetic codec while the pixels inside the arbitrarily shaped regions are coded by a Differential Pulse Code Modulation(DPCM) with uniform quantization. In the experiment, a tradeoff is found between the bitrate of the edges and the arbitrarily shaped regions through experiments which could optimize the rate-distortion performance. Experimental results show that the proposed scheme can preserve the edges well and the visual quality of the synthesized view seems better than the JPEG and JPEG2000 scheme at the high bitrate.2. A depth map coding method based on clustering of arbitrarily shaped region is proposed. In this method, at the encoder, a depth map divided into n clusters by using the K-means algorithm. Each cluster can form a new depth map, and then we can compress each of them. First, we detected a given new depth map’s edges and coded the edges with an arithmetic codec; Second, all pixels values in the same cluster except edge pixels are arranged as a row vector, and then down-sampling and entropy coding are performed on them; Finally, the bitstream is transmitted to the decoder. At the decoder, we can reconstruct each image by Partial Differential Equation; After we get all the recovered depth maps, the overall depth map is formed by integrating them together. Experimental results show that the proposed method can achieves the better quality of the synthesized views than the JPEG and JPEG2000 standard.