Research On Quality Scalable Video Coding And Error Control

With the fast development of video coding technology, network infrastructure, information and consumption electronic, the real-time streaming media applications have become one of the most promising services of information industry. However, differing from the storage oriented video applications, the streaming media applications must cope with complicated environment problems such as heterogeneous networks, fluctuation of bandwidth, transmission error and diversity of terminal devices. Thus, the video coding system is demanded to provide scalability of temporal, spatial and quality (SNR) with low computation complexity. Scalable video coding (SVC), as one of the effective solutions to deal with these problems, has theoretic and applied value, and is one of the attractive research focuses of video coding field.The work of this dissertation is concentrated on the scalable video coding techniques, based on the newest scalable extension of H.264/AVC. As to the fine granular scalability (FGS), different coding structures are studied as well as the problem of coding parameters optimization and bit-stream truncation. As to the coarse granular scalability, the error concealment of enhancement layer is researched.First, the framework of FGS and several classic improved frameworks are analyzed in detail. Aiming at the low delay video applications, a key reference frame based hybrid open-close loop FGS coding framework is presented. In the proposed framework, most frames are coded as non-key frame to achieve the best coding efficiency. The non-key frames exploit an open-structure with single prediction loop which use the highest quality image as the reference picture for both base and enhancement layer. Meanwhile, the prediction drift is introduced to base layer. To control drift, some key frames are inserted periodically. Key frames are predicted only from previous key frames, and a close-structure with double prediction loop is exploited. Thus, the drift of base layer is confined between two adjacent key frames, and the trade-off between coding efficiency and robustness is achieved. Simulation results show that, compared with AR-FGS, the proposed framework improves the coding efficiency significantly at almost whole range of bit rate, and lowers the computation complexity meanwhile. Only at the lowest bit rate points close to bit rate of base layer and to the sequences with moderate to high motion degree, the coding efficiency decreases slightly.Considering the different contributions of key and non-key frames to the proposed hybrid FGS framework, this dissertation proposes an unsymmetrical bit-stream extraction method. The available bandwidth is allocated to the key frames prior to the non-key frames among the same FGS layer. The left bandwidth is allocated among non-key frames averagely. Simulation results show that, with the proposed bit-stream extraction method, the performance of hybrid FGS framework is further improved.Second, the determination of leaky factor for AR-FGS is studied. Because the method of bit-stream extraction directly affects the performance of leaky prediction, the extraction procedure of JVT SVC test model-JSVM (joint scalable video model) is analyzed and a modified extraction method is proposed to guarantee the smoothness of extracted sub-stream on frame level. Then, an adaptive leaky factor determination algorithm is proposed. The algorithm sets the optimal leaky factor for each frame according to the ratio of current reference frame's base layer bit-rate to that of previous I-frame's. The selected factor is further adjusted according to the ratio of several previous frames'average base layer bit-rate to that of current reference frame's. Simulation results show that, over a wide range of bit-rate, the PSNR of proposed algorithm can approximate or even surpass the best performance of using fixed leaky factor.Last, this dissertation studies the error concealment of SNR enhancement layer. Through detailed analysis of the characteristics and suitable situation of two common and effective error concealment methods, a distortion estimation-based adaptive error concealment algorithm is proposed. Based on the correctly reconstructed base layer residue and the difference of quantization step between base layer and enhancement layer, the algorithm estimates the total distortions by exploiting the base layer reconstruction or the previous enhancement layer reconstruction, respectively. Thus, the error concealment method which leads to smaller distortion is selected to recover the lost area for each 4×4 block. Simulation results show that average PSNR gains of 4.0 and 0.6 dB are achieved compared with the original two error concealment methods respectively. The superiority is obvious especially for sequence with moderate activities and moderate difference of quantization step.