【作者】 周承涛；

【导师】 陈耀武；

【作者基本信息】 浙江大学 ， 电子信息技术及仪器， 2014， 博士

【摘要】 新一代视频编码标准HEVC(High Efficiency Video Coding)的编码效率比H.264/MPEG-4AVC提高了一倍以上。但是对其编码工具的灵活选择使得HEVC编码器复杂度急剧增加,这严重阻碍了HEVC的应用和发展。因此针对HEVC编码快速算法的研究至关重要。本文在介绍了视频编码技术的概况和视频编码标准的发展历史之后,对HEVC的编码框架进行了概括。然后针对HEVC编码工具的特点,指出了HEVC编码优化的方向：帧内编码单元快速选择算法、帧间编码单元快速选择算法和运动估计快速算法。针对HEVC帧内编码单元划分复杂度高的问题,提出了一种基于统计学习的帧内编码单元快速选择算法。将帧内编码单元的划分建模为k-means分类问题,通过分析四个子编码单元覆盖区域像素的方差组合而成的四维向量的特征,用简单而有效的k-means分类方法进行编码单元划分的预测,从而避免了基于率失真优化的全搜索算法,降低了编码器的计算复杂度。针对HEVC帧间编码单元划分复杂度高的问题,提出了基于深度时空相关性的帧间编码单元快速选择算法。在分析了时空相邻编码树单元之间的相关性之后,依据相关性强弱选择最佳相邻编码树单元,并利用最佳相邻编码树单元的深度,提前预判当前编码树单元的深度搜索范围。同时根据前一帧中相邻编码单元的深度关系和当前帧中已编码相邻编码单元的深度,预判当前编码单元的深度搜索范围,从而进一步提高了帧间编码单元选择的速度。HEVC中的多参考帧技术以及灵活的数据划分方式,大幅度增加了运动估计的复杂度。针对多参考帧选择,提出了基于不同预测单元最佳参考帧相关性和层间编码单元最佳参考帧相关性的多参考帧选择算法。利用划分为2N×2N的预测单元中各个参考帧的率失真代价,减少同一编码单元中其它划分模式的候选参考帧数目,加速参考帧选择过程。同时当父编码单元的模式为SKIP时,将当前编码单元中所有划分模式的参考帧限定为父编码单元的最佳参考帧,从而进一步降低多参考帧选择的复杂度。另一方面,搜索范围在影响运动搜索复杂度的同时也影响数据搬运带宽,通过分析不同分辨率视频设置不同搜索范围的编码结果,为不同分辨率的视频推荐不同的搜索范围,能够有效的降低数据搬运带宽。最后总结了本论文的研究成果,并提出了该领域下一步研究的方向和任务。更多还原

【Abstract】 The latest video coding standard High Efficiency Video Coding (HEVC) can provide the same perceptual video quality with50%bitrate reduction compared to H.264/MPEG-4AVC. However, such flexibility of encoding tools introduces great computation burden on HEVC encoder, which is the bottleneck of enabling HEVC into practical use. Hence, research on reducing the computational complexity of HEVC encoder while maintaining the high coding efficiency is crucial for HEVC.It gives a brief introduction of video coding techniques and the development of video coding standards at the beginning of this thesis, and summarizes the HEVC encoding framework. According to the features of HEVC, the key areas for fast algorithms are figured out:fast intra coding unti size selection, fast inter coding unit size decision, and fast motion estimation algorithms.To reduce the computational complexity of HEVC intra encoder, a fast intra coding unit (CU) size selection algorithm based on the statistical learning is proposed. To address this issue, the splitting of CU is modeled as a k-means classification problem. The four dimension vector is formed by using the pixel variance of the four sub-blocks in a CU and selected as the feature for the k-means learning. In such a way, the full search based on the rate distortion cost for all possible prediction units is avoides and the computational complexity is reduced.To fasten the inter CU size decision, a spatio-temporal correlation based fast CU size decision algorithm is proposed. After analyzing the correlation between the current coding tree unit (CTU) and the spatio-temporal neighbor CTU, the best spatio-temporal neighbor CTUs are selected. The depth information of the best spatio-temporal neighbor CTUs are used to predict the depth range of the current CTU. The characterstic called the depth monotonicity between the current CU and the adjacent CU is introduced. Using this characteristic, some specific depths are skipped when performing the CU size selection in the current CU. Thus, the computational complexity of the HEVC encoder can be significantly reduced. The multiple reference frame and the flexible data representation in HEVC increase the complexity of motion estimation. A fast multiple reference frame selection (MRFS) algorithm is proposed by exploiting the correlation among different prediction units (PUs) in the same CU and the correlation between the parent CU and the child CU. After performing the MRFS of2N×2N PU, the rate distortion cost of each reference frame is used to reduce the number of candidate reference frame for other PUs. When the parent CU is SKIP, the best reference frame of current CU is fixed as the best reference frame of the parent CU. Thus, the computational complexity of MRFS is reduced. To reduce the bandwidth of the movig data for motion estimation stage, the author analyzes the influence of the search range for sequences with different resolutions and characteristics. And the intial search ranges for sequences with different resolutions are suggested, and the data bandwidth can be reduced. Finally, the author concludes the new achievements of the whole research and the prospect of the future research.更多还原