Please use this identifier to cite or link to this item:
|Title:||Distributed video coding and its application to error resilient video communication||Authors:||Zhang, Yixuan.||Keywords:||DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems||Issue Date:||2013||Source:||Zhang, Y. (2013). Distributed video coding and its application to error resilient video communication. Doctoral thesis, Nanyang Technological University, Singapore.||Abstract:||In current centralized video coding (CVC) systems including all the ITU-T and MPEG standards like H.264/AVC, the encoder is designed to compress input signal by exploiting all the temporal, spatial and statistical redundancies presented in a video sequence. As a result, all the centralized video encoders are much more complex than their corresponding decoders. Such framework is suitable for broadcasting applications where compressed bitstreams are repeatedly served to users. On the other hand, according to Slepian-Wolf theorem and Wyner-Ziv theorem for lossless distributed source coding (DSC) and lossy DSC respectively, it is also possible to accomplish efficient signal compression by exploiting source correlation at the decoder only. By applying the DSC mechanism in video compression known as distributed video coding (DVC), we can shift the correlation exploitation tasks of high computational complexity from the encoder to decoder, especially that of exploiting temporal correlation in video coding, thus achieving low complexity encoding. This feature is highly desired in the emerging wireless video communication applications where compression needs to be accomplished at the encoders with the scarcity of computation and especially power resources, such as wireless video surveillance cameras, tablets and smart phone cameras, and remote video sensors. Despite the great efforts spent on the research of DVC in the last decade, the rate-distortion performance gap between DVC and CVC remains significantly large. There are three main sources for the performance gap: less efficient channel-code based distributed source coding vs. entropy source coding, less reliable side-information estimation and less accurate correlation prediction. To enhance the performance of DVC systems in terms of coding efficiency, in this thesis, we consider quantizer design with side-information and high-order correlation estimation. We first address quantizer design by studying two coset partition based quantization schemes for better reconstruction quality and lower bit rate, respectively. Meanwhile, by exploring the high-order correlation in wavelet domain to improve the correlation model accuracy, we lower the bit rate in DVC. More details are elaborated in the following. One of the practical quantizer designs in DVC is the coset partition based one-dimensional nested lattice quantizer which use a scalar quantizer followed by a coset channel code. Two typical coset partition methods, namely, modular based coset partition and scalar quantizer based coset partition, are examined in this thesis. The modular based approach introduces no reconstruction error when the correlation noise is small, but suffers a large error when the correlation noise goes beyond the error correction capability of the coset channel code. In contrast, the scalar quantization (SQ) based binning has an advantage that the maximum decoding error can be clipped in a certain range even though there is an unexpected large correlation noise, but errors may be inevitable even the correlation noise is small. Taking advantage of their respective strengths while circumventing their weaknesses, an adaptive coset partition scheme is proposed by integrating the two coset partition methods to minimize the decoding errors given a rate constraint. Different modes are designed empirically for bit allocation between the two coset partition schemes based on the correlation noise prediction. To further extend the study of quantizer design using coset partition to improve the performance, the nested quantization is investigated in a more analytical way. The adaptive nested lattice quantization scheme for distributed source coding is modified to minimize the rate for a given distortion. An analysis is presented to determine a correlation noise threshold based on which the two above mentioned coset partition schemes are employed adaptively. More specifically, we derive the overall rate of indices generated by different schemes with respect to the threshold, and then solve the problem numerically to find the threshold which minimizes the rate. The correlation modeling is another key to the performance of DVC. In the most widely used channel code (low-density-parity-check (LDPC) codes and turbo codes) based DVC, correlation modeling is used to initialize the belief propagation decoding process by providing a priori probability estimates for the bits received. A better estimation accuracy of the probability tends to improve the coding efficiency. As the existing Laplacian correlation noise modeling fails to exploit high-order statistical correlation, an effective approach is proposed to explore the inter-coefficient correlation across scales and inter-bit correlation within each frequency band in wavelet domain. Using the two levels of correlation exploitation plus the widely used Laplacian correlation noise modeling, we can achieve better a priori probability prediction through Bayesian approach. The proposed scheme is implemented in a recently developed wavelet domain DVC framework with significant and consistent coding gain achieved. Apart from enhancing the coding efficiency of the DVC systems, we also investigate the error resilient video coding based on DSC. Error propagation is a common problem in conventional video coding in case a reference symbol is corrupted. It is known in DSC, the source coding can be considered as a virtual channel coding, where the reference is a corrupted version of the current source. Therefore a unified single channel code can be employed for joint source-channel coding which combats the virtual and real channel errors simultaneously. In this way, we propose a channel-aware error resilient video compression scheme using the DSC technique to stop error propagation. We apply a single distributed source code to certain frames for both compression and data protection purposes, thus stopping error propagation more efficiently. Compared with the separate source and channel coding, the proposed joint source-channel video coding scheme shows the rate saving in eliminating error propagation.||URI:||http://hdl.handle.net/10356/54842||metadata.item.grantfulltext:||open||metadata.item.fulltext:||With Fulltext|
|Appears in Collections:||EEE Theses|
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.