【作者】 林贵旭；

【导师】 郑世宝；

【作者基本信息】 上海交通大学 ， 信号与信息处理， 2008， 博士

【摘要】 近年来,随着信息社会的发展,人们对视频信息的需求不断增长,相应的视频产品快速进入到办公、娱乐和安全等众多领域当中。在此推动下,数字视频编码、处理和传输技术取得了长足的发展。作为新一代的视频压缩编码标准,H.264在压缩效率、质量和网络适应性等方面均取得了显著的提高。目前H.264正逐渐取代原有视频压缩标准在各应用领域占据主导地位。视频标准的应用和推广也离不开相应码率控制方案的支持。由于带宽和存储资源相对有限,码率控制对于视频编码器输出码流与信道速率的匹配,以及在码率受限的条件下重建视频质量的优化都具有十分重要的意义。然而,H.264视频编码器自身具有的高度复杂性使得其码率控制相对以往的视频标准更具挑战性。本文面向实时视频应用,通过对H.264码率控制技术的研究和优化,力求提高重建视频的主观质量。本文首先简要介绍了H.264采用的若干先进编码技术,特别是与码率控制紧密联系的帧内/帧间预测技术和结合整数DCT变换的量化策略。接着对率失真理论在视频编码优化以及码率控制中的应用进行了分析,并总结出码率控制方案中常用的几个率失真模型。随后重点阐述了码率控制的几个关键问题,在对JVT推荐的系列H.264码率控制参考算法进行回顾的基础上,分析了现有H.264码率控制技术的缺陷以及对此进行优化的可能性。然后,针对现有H.264帧级比特分配方案在视频序列包含快速运动内容或者场景切换时可能导致PSNR剧降的问题,提出了一种改进的基于预编码的H.264帧级比特分配方案。算法首先对帧内所有宏块进行16×16模式的预编码,并据此计算出“码率-失真之比”以度量当前帧的编码复杂度。目标帧比特数最终由帧编码复杂度、帧间PSNR波动程度以及缓冲区状况共同决定。实验结果表明该方案能够有效地抑制由于快速运动或者场景切换而导致的PSNR剧降,并取得一定的PSNR增益。此外,由于预编码不依赖于后续编码帧,且导致的额外计算复杂度也在一个可接受的范围内,这为算法的实时应用提供了可能。为了实现质量恒定的实时视频编码,改善重建视频图像的观看效果,本文将ρ域率失真模型与二步编码框架相结合,提出了一种面向实时应用的视频质量平滑的H.264码率控制算法。算法第一步对当前帧的所有宏块进行编码,根据编码结果建立当前帧的ρ域率失真模型。同时采用一个低通滤波器估算当前帧的目标失真度,并结合GOP剩余编码比特和缓冲区状态得到当前帧的目标比特数。如果第一步得到的编码失真或者编码比特超过相应的阈值,则进行第二步编码。第二步采用ρ域码率控制算法得到帧内所有宏块的最终量化参数,并对第一个编码阶段的残差信号进行再编码。该算法在准确控制码率的同时,实现了相对平滑的视频质量输出。最后,本文分析了从客观到主观的视频质量评价标准发展过程,指出根据HVS特性进行视频编码器优化的必要性。在此基础上,重点探讨了模式选择与帧内比特分配之间的关系,以及基于视觉感知的H.264模式选择优化的意义和可行性。为了进一步提高重建视频的主观质量,本文从视频编码的角度出发,提出了一个基于视觉重要性的视频图像分析模型。根据这个模型,提出了一种基于视觉感知的H.264自适应模式选择算法。算法按照每个宏块的视觉重要性等级,自适应地调整率失真优化模式选择过程中的拉格朗日算子,使得视觉重要性等级高的宏块采用失真较小的编码模式,视觉重要性等级低的宏块采用码率较低的编码模式,从而实现了基于视觉感知的编码性能优化。针对码率受限的应用环境,本文在自适应模式选择算法的基础上进一步提出了一种宏块级的码率控制策略,有效提高了重建视频的主观质量。更多还原

【Abstract】 In recent years, with the development of information society,the demands for multimedia increase largely. Corresponding digital video products are popular in broad application areas such as business, amusement and surveillance, etc. This drives the digital video techniques of encoding, processing and transmission to advance rapidly. Being the newest international video coding standard, H.264/AVC achieves great improvements in coding efficiency, quality and network adaptability. As a result, H.264 gradually becomes dominant in video applications.Meanwhile, rate control plays an important role in the application of video coding standards. For a video encoder, rate control is employed to maximize the coding quality at certain channel bandwidth and buffer constraints. However, the interdependence of rate control and RDO in H.264 makes it more difficult to implement than that in prior standards. In this thesis, we focus on improving the perceptual quality of real time video applications by optimization of H.264 rate control scheme.Firstly, the advanced coding features of H.264 are introduced, especially for the Intra/Inter prediction methods and novel quantization scheme combined with interger DCT-like transform, which are closely related with rate control. Then we give an analysis of rate distortion theory in video encoder and rate control optimization. The classical rate distortion models for rate control are reviewed. Subsequently, we sum up some key issues in rate control. Based on the JVT adopted rate control schemes, we give a discussion on the shortage of existing H.264 rate control algorithms and the probabilities of improvements. To relieve the video quality degradation caused by high motions and scene changes, an improved frame-layer bit allocation scheme based on pre-encoding is proposed. First, the frame is pre-encoded in 16×16 modes with a fixed quantization parameter (QP). The coding complexity of current frame is then measured by rate-distortion ratio according to the coding results of pre-encoding stage. The frame bit budget is finally decided by the frame coding complexity and inter-frame PSNR fluctuation, combined with the buffer status. Simulation results show that, in comparison with the H.264 adopted rate control scheme, our method is more efficient to suppress the sharp PSNR drops due to high motions and scene changes. The pre-encoding stage introduces no more than 30% extra computational complexities and need no any following frame, which makes it applicable to real time video coding.To futher improve the visual quality of real time video applications by means of smoothing the frame quality fluctuations, we develop a two-stage rate control scheme. It integrates theρ-domain R-D model into a two-stage H.264 coding framework. In the first stage, a coarse QP is used for motion estimation and RDO. Based on the coding results of this stage, the relatedρ-domain R-D model is constructed. Then a low-pass filter is used to estimate the target frame distortion. The frame bit allocation is determined by target distortion, remaining coding bits and buffer status. If the bit rate or distortion of the first stage exceeds related thresholds, the second coding stage is performed, which adopts theρ-domain rate control scheme to get the refined QP of each MB for residual signal recoding. Experimental results show the proposed scheme can efficiently smooth out the vidual quality fluctuation between successive frames, on the basis of meeting bit rate constraints accurately.Finally, the development of video quality measurement is reviewed, based on which we give a discussion on the necessities of video encoder optimization in terms of HVS characteristics. And the relationship between mode decision and bit allocation within a frame is analyzed. Then we focus on the benefits and feasibilities of perceptual optimization for H.264 mode decision. According to above analysis, a perceptual importance analysis model is developed for video segmentation from the view of video encoding. Base on this model, we propose a perceptual mode decision scheme for H.264, which adaptively adjusts the Lagrangian multiplier in RDO process in terms of the perceptual importance of each MB. Therefore, in a frame, more bits can be allocated to the regions with high perceptual importance for visual quality improvement and few bits to the regions with low perceptual importance for bit rate saving. To meet bit rate constraints, we proposed a MB-layer rate control algorithm on the basis of adaptive mode decision scheme. The simulation results show the efficiency of the proposed scheme.更多还原