Multiple Description Coding : Proposed Methods And Video Application

Multiple description (MD) coding and quantization has received a lot of attention recently, and has been studied extensively and extended to many demanding applications such as speech and video. We propose two multiple description quantization schemes in order to design the codebooks and partitions of side and central quantizers. The applied framework originates in the multiple description quantization via Gram-Schmidt orthogonalization approach which provides systematic treatment of the achievable rate-distortion region by subtractive dithering and successive quantization along with quantization splitting. The basic idea of our proposed MD quantization schemes is to minimize a Lagrangian cost function (defined as the weighted sum of the central and side distortions) by an iterative technique which jointly designs side codebooks and consequently forms associated partitions. In the first scheme, multiple description vector quantization with weighted sum central decoder (MDVQ-WSC), the central decoder is formed by a linear combination (weighted sum) of the side codebooks. The parameters of this linear combination are also found to minimize the central distortion. Once the side codebooks are found by the iterative technique at the final iteration, we propose to replace weighted-sum central decoder with the optimal decoder to enhance the performance of central decoder and achieve lower central distortion. In the second scheme, multiple description vector quantization with i optimum central decoder (MDVQ-OC), the central codebook is found by the optimal decoder. This increases the complexity of the scheme, but also achieves lower central distortion. We compare the performance of our proposed methods with the optimum MD quantization and also with the optimal rate-distortion bound for the case of a memoryless Gaussian source. We demonstrate by simulations that our proposed methods perform very closely to the optimum MD quantizer with considerably less complexity and with a few iterations. We also propose a multiple description video coding technique motivated by human visual perception in order to generate two correlated streams. We first calculate two simple parameters to characterize the smoothness and edge features of each block of an MPEG video frame. We employ these two parameters as a measure of the perceptual tolerance of DCT blocks against visual distortion. We duplicate the key information such as motion vectors and some low-frequency DCT coefficients, and split the remaining DCT coefficients of prediction errors according to the calculated perceptual tolerance parameter. The implementation of the proposed MD video coding preserves the core of the standard video coder and does not significantly increase the coding complexity. Simulation results reveal that our simple MD video coding method achieves superior performance compared to other similar techniques, such as simple rate-splitting method, which lack to address perceptual distortion in the design problem.