Rate-distortion Optimized Multiple Description Coding With Applications To Image And Video Communication

MDC (Multiple description coding) refers to encoding the source into severaldescriptions. Each description can be decoded separately with acceptable quality.When more than one descriptions are received, joint decoding can be performedfor better quality. First proposed decades ago, MDC is both of theoretical interestand practical relevance. On one hand, MDC, as a basic problem in informationtheory, is strongly related to many other theoretical problems. On the otherhand, MDC has been regarded as a promising approach to combat packet loss fornetwork information transmission. Since sources of diferent kinds usually havediferent statistical characteristics, when designing MDC for a particular kindof sources, we need to integrate the principle of MDC with their characteristics.Based on the theory of MDC, we study how to design and optimize MDC schemesfor discrete sources, images, and videos, respectively, in order to improve the R-D(Rate-Distortion) performance of practical multiple description codecs.We frst study MDC algorithms for discrete sources. Past research in thisrespect mostly focuses on the case in which the source statistics are known apriori, while in practice source statistics are usually unknown or time-varying.In this thesis, UMDC (Universal Multiple Description Coding) is studied, whichrefers to performing MDC without knowing the source statistics in advance andgradually improving the R-D performance by adapting the coding parameters.We generalize the principle of natural type selection, which is originally proposedby Zamir and Rose for universal single description coding, to the setting of MDC.We propose two UMDC schemes based on random codebooks, with one of thembeing fxed-rate and the other being fxed-weight. The R-D functions of both schemes are derived, which coincide with the EGC (El-Gamal-Cover) bound ifthe coding parameters are optimized. For both schemes, we show that the jointtype of reconstruction codewords can be used improve their R-D performance.We then propose a practical UMDC scheme for binary sources based on ourtheoretical results and contemporary MDC methods. Experiment results showthe efectiveness of our scheme.Since a variety of schemes have been proposed for multiple description imagecoding, we study how to integrate the advantages of various existing schemes. Wepropose a combinatorial-optimization-based MDC framework for vector sources.In this framework, the high-dimensional source vector is frst partitioned intoseveral mutually disjoint tuples of equal length, after which each tuple is codedwith a low-dimensional scheme. Given the R-D functions of the low-dimensionalschemes, we show that the R-D optimization of the whole framework reduces toa combinatorial optimization problem, which in certain cases admits polynomial-time solutions. We propose a R-D optimization algorithm for the whole frame-work and apply it to multiple description image coding. Experiment resultsshow that our approach can help to improve the R-D performance of multipledescription image coding.Recently, stergaard and Zamir proposed a MDC method based on DSQ(Delta-Sigma (Σ-△) Quantization). Although in the quadratic Gaussian casethis approach achieves the theoretical bound, research of its application to imagecoding is still at starting stage. In this thesis, we propose a vector-DSQ-basedmultiple description image codec. The source image is frst transformed intoa block sequence, after which vector DSQ is performed with a bank of noise-shaping flters. R-D optimization is used to select the flter coefcients andquantization steps, which also alleviates the zero-to-nonzero fipping problemof DSQ-based image coding. A post-processing algorithm is proposed to boostthe reconstruction quality of the side decoders. Experiment results show thatthe proposed scheme achieves improvement in terms of both PSNR values andsubjective quality.Diferent from images, it is one of the key problems in designing MDVC (Multiple Description Video Coding) schemes how to deal with the situationwhen the reference frame at the decoder difers from that at the encoder. Theproblem is called drift problem in the literature, which might cause severe videoquality degradation. We use the method of distributed source coding (Wyner-Zivcoding) to alleviate the drift problem of MDVC on packet loss channels. We frstpresent an asymptotically optimal code design of MDWZ (Multiple Descriptionsin the Wyner-Ziv setting). Then, a two-channel DMDVC (Distributed MultipleDescription Video Coding) scheme is proposed. The scheme performs MDWZcoding on each inter-coded frame. Side informations can be interchanged betweenthe side decoders without loss of decoding quality. When the side informationof a side decoder is damaged by packet loss, it can use the side information ofthe other side decoder, while in prediction-based MDVC, drift occurs when sideinformations are interchanged. Experiment results show that DMDVC is robustat medium-to-large packet loss rates.