【作者】 李志成；

【导师】 顾畹仪；

【作者基本信息】 北京邮电大学 ， 电磁场与微波技术， 2009， 博士

【摘要】 视频通信技术的发展,是与其承载网的发展分不开的。视频通信业务可以承载在不同的底层网络上,从早期的PSTN(Public SwitchedTelephone Network,公共交换电话网),到窄带ISDN(Intergrated ServicesDigital Network,综合业务数字网),ATM(Asynchronous Transfer Mode,异步传输模式)和现在的IP网络,都可以成为视频通信业务的承载网。目前,随着底层光网络技术和IP技术的融合,IP数据光网络的概念被提了出来,它被理解为一种能够提供IP数据服务的、以光网络为底层传送网络的宽泛的网络概念。IP数据光网络可以表示为IP/…/Optics的形式。基于IP数据光网络的视频通信技术也成为了一个重要的研究方向。目前,无论是服务提供商还是用户,对于视频通信技术的期望都是能够获得有QoS(Quality Of Service,服务质量)保证的视频服务。而对于视频通信质量具有严重影响的因素之一,就是网络的丢包。因此,更准确的理解网络丢包特性与视频通信质量的关系,就显得尤为迫切。由丢包引起的视频通信质量下降这一问题一直是工业界和学术界研究的热点。对于承载于IP数据光网络的视频通信而言,我们同样关注端到端的网络丢包对视频通信质量的影响。研究基于IP数据光网络中端到端丢包与视频通信质量之间的关系,其中最关键的就是要建立适合IP数据光网络丢包特性的视频传输失真模型。为视频传输失真建模的主要目的,就是希望可以准确预测由丢包引发的视频解码失真。在给定网络丢包信息的情况下,如何准确地估计出由丢包产生的视频传输失真,或者说如何建立准确的视频传输失真模型,对于更好的理解IP数据光网络端到端丢包与视频质量之间的关系具有重要意义。然而,由于IP数据光网络中端到端的丢包是具有记忆性的突发式丢包过程,其丢包特性可以由马尔可夫丢包模型来描述,而以往的视频失真模型都是针对贝努利丢包网络的,所以,为IP数据光网络建立视频通信失真模型就更为复杂。本论文针对上述问题进行了深入研究,获得了若干具有创新性的成果,根据其中部分成果写成的论文已经被IEEE会刊IEEE Transaction on Circuits and Systems for Video Technology录用。论文主要的工作和创新点包括以下几个方面:第一,为了研究IP数据光网络中具有突发特性的端到端丢包对视频通信质量的影响问题,首先将IP数据光网络中的端到端丢包纳入到具有普适性的(m+1)状态马尔可夫丢包模型中,并在这种丢包模型下,提出了视频通信失真的统一的研究框架;第二,要研究IP数据光网络中这种马尔可夫式的丢包对视频失真的影响,首先就要建立可以计算任意丢包模式造成的视频失真的数学模型。针对采用混合式编码的视频通信,提出了适用于任意丢包模式下的视频通信失真模型“失真传递模型(Distortion Infection Model)”,该模型复杂性低,比之前的传统模型更加精确;第三,基于失真传递模型,提出了二状态马尔可夫丢包过程,即吉尔伯特丢包过程下的视频失真模型“失真网格模型(Distortion TrellisModel)”,该模型可以计算给定吉尔伯特丢包过程导致的视频通信的MSE(Mean Square Error,均方误差)失真期望;第四,由于原始的失真网格模型复杂度高,计算量大,因此提出了快速算法“滑动窗口算法(Sliding Window Algorithm)”,使在精度损失很小的情况下模型的计算量降低90%以上。基于该算法,失真网格模型更加实用化,而且可以满足实时化要求。第五,对失真网格模型进行扩展,使其可以计算(m+1)状态马尔可夫丢包过程下的视频通信失真,并结合实验网络中采集到的数据说明了模型的具体使用方法。基于本文提出的数学模型,详细分析了IP数据光网络中端到端的丢包过程各参数对视频通信质量的影响,观察到了一些之前未见于其他文献的结果,并通过所提出的数学模型进行了解释,为今后的差错控制技术提供依据和理论基础。文章最后对本文提出的数学模型一些可能的应用做了简要的介绍。更多还原

【Abstract】 The development of video communication is along with and partly based on the advancement of its underlying network technologies. Generally, video services could be carried through various networks, including the early PSTN, ISDN, and later ATM, as well as the IP networks which are widely deployed currently. Today, a new concept of underlying networks, IP data optical network, has been proposed, to present a relatively broad category of networks, which are based on optical techniques and enable the IP services. It could be denoted as IP/.../optics for clarity. In such a case, researchers have turned into the new subject of video over IP data optical networks.The QoS (Quality of Service) in video transmission over IP data optical networks is most important for both service provider and end user. The key factor degrading the video quality in video communication systems is the packet loss. It is thus important to deeply understand the relation between the end-to-end packet losses and the user-perceptive video quality. That is why people in both institutes and industry are interested in this subject for a long time. For video over IP data optical networks, we also focus on the same problem, i.e. the impact of IP data optical network end-to-end packet losses in user-perceptive video quality. To understand the effect of packet loss on video quality, it is desired to model the end-to-end distortion caused by packet loss in decoded video. Based on the distortion model, one can estimate the packet-loss-induced distortion at the encoder for video transmission over lossy channels. It is thus critical for most joint source-channel rate-distortion optimized schemes and channel error control techniques such as inter/intra mode switching and forward error correction.When modeling the packet-loss-induced distortion for IP data optical networks, the network loss characteristics should be prior known. A simple assumption is to regard the packet losses as an independent and identically-distributed random process, characterized by a Bernoulli loss model. However, for IP data optical networks, the end-to-end packet losses often exhibit time dependences. It will lead to burst packet loss, a characteristic cannot be found in a Bernoulli loss model, but can be described by a Markov loss model. Many distortion models for video over lossy networks have been proposed. However, all existing models are based on the Bernoulli loss assumption. Modeling the video decoded distortion for Markov-model losses is more complicated than that for Bernoulli losses. This paper focuses on the distortion modeling problems and aims to made one step effort on understanding the impact of IP data optical network losses on user-perceptive video quality. Part of the results proposed in this paper will be published in the IEEE Transaction on Circuits and Systems for Video Technology.The main contributions of this paper are summarized as follows.(1) To formulate the problem in a mathematic way, we first propose a framework of video decoded distortion modeling, where the packet losses in IP data optical networks are modeled as an (m+1)-state Markov chain.(2) To establish the distortion model for video transmission over IP data optical nerworks modeled by Markov loss process, the decoded distortion for arbitrary packet loss pattern should be prior modeled. For video communications using motion composition techniques, we propose a distortion model, denoted as the Distortion Infection, to estimate the decoded distortion caused by arbitrary packet loss pattern.(3) Based on the detailed analysis of both the error propagation and the loss burstiness, the Distortion Trellis model is established, enabling us to estimate the expected MSE distortion for two-state Markov losses, or Gilbert losses, at both frame level and sequence level at the encoder. The model is designed to be applicable to most block-based motion compensated encoders. The model also allows for any temporal error concealment at the decoder.(4) A siding window algorithm is developed to calculate the MSE estimation with low complexity. Using the sliding window algorithm, in most cases more than 90% computation burden can be saved compared with the original Distortion Trellis model without degrading of accuracy.(5) Finally, we extend the proposed model to a more general form, enabling us to calculate the distortion caused by (m+1)-state Markov losses, and thus to estimate the decoded distortion for video transmission over IP data optical networks. Based on the proposed model, we analysis in detail the impact of IP data optical networks losses on video quality and established some new findings that have never been proposed before. The last section mentions some probable practical applications of the proposed models.更多还原