【作者】 沈礼权；

【导师】 张兆扬； 王国中；

【作者基本信息】 上海大学 ， 信息与通信工程， 2008， 博士

【摘要】 H.264是ITU-T的VCG和ISO的MPEG最近联合开发的新一代视频编码标准。H.264出色的编码效率和有效的传输接口使得它可以广泛应用于IP网络和无线网络上的视频业务。H.264存在着与早期的视频编码标准相似地研究课题。首先,编码效率在比以前标准提高50%的同时,复杂度也增加了一个数量级。如何减少计算复杂度对H.264标准的实时应用尤为重要。其次,H.264码率控制方案假设视频序列的帧与帧之间有一定的相关性,可以用前面的编码产生的统计信息指导当前帧的码率控制。然而,视频序列帧与帧之间统计特性千差万别,尤其在场景切换或者运动剧烈的情况时,前后帧之间的相关性消失或明显减弱。这造成了序列前后编码帧的图像质量有明显的波动。如何改进码率控制方案以降低前后帧解码图像的波动对H.264的应用也尤为重要。最后,H.264在提高50%编码效率的同时,编码复杂度也增加了十几倍。单纯从基于混合编码框架上来提高编码效率已越来越困难。近年来,国内外学者逐渐加大了将人眼视觉特性应用到视频编码中的研究力度,以在进一步提高编码效率的同时,改善图像的主观质量。如何将人眼视觉特性融合于编码系统中,将是视频编码技术发展的一个重要问题。为此,本文将从降低编码复杂度、码率控制、视觉编码这三方面进行研究以优化H.264的编码算法。第一,基于H.264实时应用,本文对H.264编码算法进行优化。针对编码过程复杂度高的整数变换与量化、运动估计、预测模式选择、参考帧选择等模块进行深入的研究与分析,首先在像素域内充分利用纹理特性、相邻块的相关性以及已编码块的信息等,给出了一套快速算法,有效地减少编码器的运算量。接着从运动场的空域、时域分布的角度出发,基于帧内模式选择、帧间模式选择,参考帧选择与运动场空时一致的相关性,给出了一套快速算法,这套算法克服了在像素域利用纹理特性的快速编码算法的局限性。第二,基于帧层、GOP层码率分配存在的局限性,本文对H.264码率控制提出了改进。给出了一种基于增量式PID和帧复杂度的帧层目标比特分配方案。首先,引入增量式PID对缓冲区“液面”进行控制,减小由于缓冲区“液面”波动带来对帧层目标比特分配的负面影响。接着,给出了基于帧复杂度的目标比特分配方案,其中采用前一帧MAD、前一帧时域活动度以及当前帧时域活动度来来预测当前帧的复杂度。本方案可使视频质量PSNR的波动平均降幅32%;给出了一种基于增量式PID算法GOP层目标比特分配方案。给出的GOP层目标比特分配方案使PSNR的波动的平均降幅达15%。第三,本文研究了如何将人眼视觉关注度特性引入到视频编码中,以在进一步提高编码效率的同时,改善图像的主观质量。提出了一种基于人眼视觉关注度的视频编码方法。该方法首先快速分割局部运动宏块、边缘宏块、平滑宏块和纹理宏块。接着,基于人眼视觉关注度高低,对不同区域宏块的量化参数进行调整。本方法在提高视频主观质量的同时,可降低10%左右的码率:提出了一种基于人眼视觉关注度的H.264码率控制算法。算法首先提出了基于运动关注度帧层目标比特分配方案,增加运动剧烈帧分配的比特数,降低运动平缓帧分配的比特数。接着,提出了基于人眼视觉关注度的宏块层目标比特分配,增加关注区域宏块的目标比特数,降低非关注以及能够掩盖更多失真区域宏块的目标比特数。本方法在降低视频质量波动的同时,关注区域的PSNR也有大幅提高,从而有效的改善了视频图像的主观质量。更多还原

【Abstract】 Compared with previous coding standards such as H.26x and MPEG, H.264 can achieve 50% bit rate reduction with the similar video quality. Excellent coding efficiency comes from adopting additional new tools such as variable size motion estimation, multiple reference frames and so on. However, these new tools also lead to extremely high computation and limit the use of H.264 in real-time applications. Thus, how to reduce the codec computation is an important aspect. H.264 rate control scheme assumes that the frame complexity varies gradually from frame to frame. However, this assumption of similar variation itself may not hold, especially for videos with high motion or frequent scene cuts. It may introduce a large distortion variation sensitive to human over frames. Thus, how to overcome the above limitation is an important aspect in implementation of a H.264 encoder. When H.264 achieves 50% increase code efficiency, the complexity also increases heavily. It is very difficult to increase more efficiency only under the traditional hybrid video coding scheme. Recently, technologies with consideration of human vision system (HVS) characteristic are applied in coding process to increase efficiency. Thus, how to efficiently eliminate the perceptual redundancy in video sequences, and how to incorporate the perceptual cues into the coding process are key issues of the future video coding. The research of this thesis will focus on these three aspects.To achieve maximal encoding speed, we analyze the most time-consuming operation such as integer transformation, motion estimation, mode decision and reference frame selection. Based on the correlation among the neighboring blocks, information of motion estimation from previous searched frames and texture information, we propose several fast algorithms as follows. After the analysis of quantization and transform operations, we first propose an enhanced all-zero block detection algorithm. We also propose an adaptive and fast fractional pixel search algorithm to reduce the computation load of motion estimation. The fractional search algorithm adopting the proposed bypass strategy and the early terminated strategy greatly improves the Center Biased Fractional Pixel Search method. A new fast intra prediction algorithm is proposed to improve the intra mode decision using three criteria, i.e., mode correlation of neighboring MBs, early termination based on all-zero block detection, and the prediction of intra mode size using texture feature of MB. A novel inter mode decision algorithm utilizing SAD of each 4×4 block from the initial test of inter 16×16 mode and texture characteristic is proposed to reduce the candidate mode set. In this thesis, we also propose an adaptive and fast multi-frame selection algorithm that uses the correlation among the neighboring blocks and information of motion estimation from previous searched reference frames.Recently, most H.264 fast algorithms only exploit the texture information in the spatial domain. They suppose that the texture homogeneous region tends to have similar motion while the region with complex texture has disorder motion. However, it does not always hold true in general video sequences since the region with complex texture may exhibit homogenous motion in some cases. In fact, the selected mode and reference frame have a large correlation with motion characteristic. MBs in the region with relatively static or smooth motion are more likely to be coded using larger block size (16×16) and the nearest reference frame, while for MBs in the region with complex motion or motion edges, more block sizes and reference frames should be evaluated. Thus, in this thesis, we propose a fast inter mode decision and a fast multiple-frame selection algorithm utilizing the spatial and temporal continuity of motion field. Experimental results show that the proposed algorithms can achieve remarkable computation cost saving and have a consistent gain for all video sequences compared with the methods using the texture homogeneous.To overcome the limitation of frame-level bit allocation, we improve H.264 rate control scheme using two tools, the incremental proportional-integral-differential (PID) algorithm and the frame complexity estimation. The incremental PID algorithm is first introduced to control the buffer and reduce the influence of the buffer abrupt fluctuation in the process of frame-level bit allocation. To reduce more video quality variations, the frame target bit allocation is also adjusted by the frame complexity from the residual energy. Experimental results show that the proposed rate control scheme decreases the average variations of video quality by 32%. To overcome the limitation of GOP-level bit allocation, we propose a novel GOP-level bit allocation based on incremental PID algorithm. Extensive experimental results show that the proposed scheme, without expensive computational complexity added, decreases the average video quality variations by 15%.Finally, we analyze the characteristics of human visual system (HVS), and introduce visual attention in the video coding process. Adaptive quantization method is proposed based on the visual interest effect. The key idea of this method is to make use of concept of visual attention and assign fewer bits to regions that are far from visually fixated location. In this thesis, we also propose a novel rate control algorithm based on visual attention. The bits allocated to each frame is proportional to the local motion attention in it, more bits are allocated to a frame if the local motion in it is stronger. More bits are assigned to visual attention MBs and fewer bits to visually less important MBs. Experiment results show that our method improve the visual quality in those frames with strong local activity and reduces temporal PSNR fluctuations across frames up to 12%. Meanwhile, the subjective quality is improved in terms of the PSNR increase in visual attention region, which is large up to 1.45dB compared with the traditional H.264 rate control scheme.更多还原