【作者】 刘英哲；

【导师】 王进祥；

【作者基本信息】 哈尔滨工业大学 ， 微电子学与固体电子学， 2013， 博士

【摘要】 随着数字化进程的发展，人们对视频图像的质量提出了更高的要求，高清清晰度和高帧频视频图像广泛的应用在数字电视，视频监控和网络广播等领域。近年来，随着移动通信技术和多媒体新业务的发展，人们越来越希望在移动终端上也能进行高质量的实时视频通信。视频编码技术的发展解决了高清视频图像网络传输带宽受限的问题。但是，高清视频编码需要处理非常庞大的数据量，这给系统资源，尤其是电池资源非常有限的移动终端系统带来了严峻的挑战。并且由于市场竞争的加剧，即使对于数字电视和安防监控这样的产品，各设计公司也在不断的推出新的解决方案，以降低其产品的实现代价，增强竞争能力。H.264/AVC编码标准是目前网络视频的主要压缩技术之一，具有较高的压缩效率和编码质量，同时其编码计算量也非常巨大。本文以现有编码技术为基础，研究面向H.264/AVC的高效编码技术，以期能够进一步降低编码过程的计算复杂度，论文主要的工作包括：对帧内编码加速技术进行了研究。提高帧内编码效率的方法主要有编码模式选择算法和预测模式选择算法。编码模式选择算法将像素作为基本单位计算宏块的平坦度特性，但是逐一统计宏块内各像素的分布情况计算量较大。针对于此本文提出了一种基于子块均值分布的编码模式选择算法，该算法将44子块均值作为基本单位来统计宏块的平坦度特征。编码结果表明，和其它算法相比较，本文算法在取得较高编码性能的前提下，有效的减少了平坦度统计过程中的计算量。使用SAD作为预测模式预判标准时，具有结构规整，易于硬件实现的特点。但是SAD需要计算子块内的全部像素，并且需要提前计算出各模式对应的预测像素值，增加了SAD计算过程的计算量。本文提出了一种基于PSAD的快速预测模式选择算法。该算法定义了各预测模式的重要像素，利用各模式间重要像素的差异将其分组进行比较，简化了SAD的计算过程，避免了预测像素的计算。编码结果和硬件实现结果表明，相比于其它算法，本文算法具有更少的预测模式使用数量和更低的硬件实现代价。研究了帧间编码加速技术。帧间编码模式和宏块运动特性之间具有一定的对应关系，利用编码块运动特性的时空相关性能够获得较好的编码速度提升。但是，对于复杂运动的视频序列，时空相邻宏块间的相关性较低，进而导致算法的性能下降。为了提高宏块运动特性的预测准确度，本文提出了一种基于参考区域运动复杂度的快速帧间模式选择算法。该算法通过定义的参考区域来表示宏块的运动特性，建立参考区域运动复杂度和各预测模式的对应关系。编码测试结果表明，和H.264/AVC标准算法相比较，本文算法在PSNR和码率基本保持不变的情况下，能够平均节约70%的编码时间。和其它算法相比较，尤其是对于运动复杂的视频序列，本文算法能够获得更高的编码速度提升增益。且硬件实现结果表明，本文算法过程简单易于实现。对运动估计中的搜索范围调整技术进行了研究。目前的搜索范围调整算法多按照一定的准则从预先设定的几种固定范围内进行选取，因此算法的自适应性能不高。由于搜索范围的大小主要是由预测运动向量PMV所决定的，因此基于运动向量差值概率密度函数的算法取得了较高的运动估计性能提升，但概率密度函数计算过程复杂。针对于此，本文提出了一种基于PMV准确度的搜索范围调整算法。该算法定义了一种简单的PMV准确度衡量方法，利用PMV准确度和搜索范围间的关系动态调整搜索范围。本文还对编码区域内的运动向量进行了分析，利用编码区域运动方向一致的特性进一步减小了搜索区域的大小。从搜索点数量上看，本文算法平均搜索点数量仅为全搜索算法的2.15%；从节约运动估计时间上看，本文能够较UMHexagonS算法平均节约15%左右的编码时间；且相比于其它同类算法，本文算法具有更少的搜索点数量。研究了多参考帧运动估计中参考帧选择技术。自顶向下相关性的参考帧预测算法能够具体给出各分割模式的候选参考帧数量，因此具有较高的编码效率。但对于运动复杂的测试序列，其预测准确度不高，且预测出的参考帧数量也较多。本文对不同分割模式间的参考帧相关性进行了分析，得出基于底层分割块的参考帧选择算法具有更高的准确度和更少的参考帧使用数量。进而提出了一种基于底层分割块的参考帧选择算法，该算法使用底层8×8编码块作为基本单元，根据各分割模式之间的关系生成各模式的候选参考帧集合。编码结果表明，本文算法具有较高的预测准确度，当参考帧设置为5时，平均参考帧使用数量仅为1.25个，相比于其它典型算法，本文算法具有更少的参考帧使用数量。更多还原

【Abstract】 With the development of digitalization process, people put forward higher requirements on the quality of video. The HD resolution and high frame rate videos are widely used in digital TV, video monitoring and network broadcasting. In recent years, with the development of new multimedia business and mobile communication technology, people are eager to realize the real-time HD video communication on the mobile terminal. The video coding technology has solved the problem of limited transmission bandwidth. However, the processing data increases dramatically when coding the HD video, and also presenting a great challenge for the system with limited source, especially the mobile terminal with limited battery. And, because keen competition in the industry, even for the digital TV and security monitoring, the design companies are constantly introducing new solutions to reduce the implementation cost, and enhance competitiveness.H.264/AVC is the main coding format standard used in Blu-ray DVD and IPTV, which has higher compression efficiency and huge computation complexity. Based on the existing coding technology, we focus on the efficient coding technology to further reduce the computational complexity of the encoding process. The main works and achievements are as follows:(1) Research on the intra coding process. The coding mode selection algorithm and prediction mode selection algorithm are two ways to improve intra coding efficiency. In coding mode selection algorithm, the pixels are used as basic unit to compute the flatness characteristic of coding block, but large amount of calculations are needed in this process. Accordingly, a fats intra-coding mode decision algorithm is proposed, which is based on the distribution of the mean value on44sub-blocks. When compared with other algorithms, this algorithm can effectively reduce the amount of computation of the flatness characteristic process. The SAD has a regular structure and can be easily implement by hardware. But all pixels in the block and predicted pixels of each mode are needed to calculate, which result in excess calculation. Accordingly, a fast prediction mode selection algorithm based on the partial SAD is also proposed. The key pixels of each prediction mode are defined, and the prediction modes are categorized into several groups according with the difference of key pixels. Base on this difference, the calculation process of SAD can be simplified, and the calculation of predicted pixels also can be avoided. The coding results and hardware implementation show that, compared with other algorithms, this algorithm has less prediction mode and lower hardware costs. (2) Research on the inter coding process. The coding speed of inter prediction can be accelerate by using motion correlation of spatial and temporal neighbor macro blocks. However, in the video sequences with complex motion activity, the correlations between the adjacent macro blocks became weakly, and then lead to a decline in the acceleration of the coding speed. To accurately measure the motion complexity of MB, a fast inter-prediction mode decision algorithm is presented, which is based on the moving complexity in the reference region. The reference region is used to evaluate the motion characteristic, and the coding modes are classified into several categories according to the movement complexity. Extensive test results show that the proposed algorithm can reduce the overall encoding time by70%on average compared with the original algorithm in the H.264/AVC standard codec. Compared with other algorithms, especially for video sequences motion complexity, this algorithm can obtain a higher coding speed. The hardware implementation also results show that the algorithm is simple and easy to implement.(3) Research on the search range adjustment (SR) algorithm. At present, only several fixed predetermined SR are used in most search range adjustment algorithm, so the adaptive ability of these algorithms is not high. Because the search range is mainly determined by the predicted motion vector (PMV), the algorithm which is based on the probability density function of motion vector difference can obtain better motion estimation performance, but the calculation process of it is too complex. According to this, this paper presents search range adjustment algorithm which is based on the accuracy of PMV. In this algorithm, a simple measurement is defined to evaluate the accuracy the PMV, and SR is adaptively adjusted by using the relationship between the SR and the accuracy of PMV. The motion vectors of the coding region are also analyzed, the consistency of motion vectors are used to further reduce the SR size. Test results show, the number of search points can be reduced to2.15%when using our algorithm, and compared to the UMHexagonS algorithm,15%motion estimation time can be saved.(4) Research on search the multiple reference frame decision algorithm. The top-down prediction method can give the number of candidate reference frame, so it can give higher coding efficiency. But for the test sequence with complex motion, prediction accuracy is not high, and the number of reference frame is also increasing. In this paper, the correlations between the reference frames of different block partitions are analyzed. And the conclusion are drawn that the higher accuracy and less reference frames can be gotten when using the bottom block partitions. Then a fast multiframe selection algorithm is proposed, which is based on the88block partition. According to the relationship between the different partitions, the reference frames of each partition are predicted by using the88partition. Experimental results proved, when applying our algorithm, the reference frame numbers can be reduced from5to1.25. Compared to other typical algorithms, our algorithm obtains a much reduction in the coding time, with the same bit rate and PSNR.更多还原