HEVC帧内预测关键技术并行算法的设计与实现

发布时间：2018-02-20 08:13

本文关键词： HEVC CUDA 并行算法帧内预测　出处：《大连理工大学》2015年硕士论文　论文类型：学位论文

【摘要】：随着人们对视频质量要求的不断提高,在H.264/AVC视频编码标准之后于2013年,国际视频编码专家组VCEG和动态图像专家组MPEG联合推出了最新的高性能视频编码标准HEVC, HEVC的目标是比H.264节省大约50%的码率。HEVC编码性能的提高是以计算复杂度大幅度增加为代价的,为了更广泛的应用HEVC,有必要提高编码速度,于是如何提高HEVC编码效率成为了研究热点。目前大多数计算机中最主要的两种处理器分别是多核CPU和众核GPU, GPU拥有大量的运算单元,适用于通用并行计算。NVIDIA公司推出的计算机统一设备架构——CUDA为GPU编程提供了很好的平台。本文基于HEVC标准算法,针对帧内预测中的关键技术,设计相应的并行算法,并基于CUDA进行实现。文中首先分析数据之间的相关性,然后针对亮度分量和色度分量帧内预测求取预测值设计不同的并行算法；针对整数DCT变换和反变换,基于蝶形快速算法设计并行算法；针对量化和反量化设计并行算法；针对帧内预测并行算法进行优化。另外,本文设计了一种逐步缩小模式搜索范围的帧内预测快速算法,在预测精度损失较小的条件下将其计算量减小一半以上。针对快速算法设计了并行算法,并基十CUDA实现。本文对设计的各个并行算法均在CPU+GPU异构平台上采用CUDA语言进行编程实现,并使用高清视频序列进行了大量实验。实验表明,本文的帧内预测并行算法相比于原始算法在保证图像质量的前提下,加速比可达5.6倍；快速帧内预测并行算法相比于原始算法在基本不改变图像质量的前提下,加速比可达8.5倍。
[Abstract]:In 2013, after the H.264 / AVC video coding standard, with increasing demand for video quality, The international video coding expert group VCEG and the dynamic image expert group MPEG jointly launched the latest high-performance video coding standard HEVC. The goal of HEVC is to save about 50% bit rate compared with H.264. The improvement of the performance of HEVC coding is at the cost of a significant increase in computational complexity. In order to apply HEVC more widely, it is necessary to improve the coding speed, so how to improve the efficiency of HEVC coding has become a research hotspot. At present, the two main processors in most computers are multi-core CPU and multi-core GPU, and GPU has a large number of computing units. The computer unified device architecture, which is suitable for general parallel computing. NVIDIA, provides a good platform for GPU programming. Based on the HEVC standard algorithm, this paper designs the corresponding parallel algorithm for the key technology of intra prediction. In this paper, we first analyze the correlation between data, then design different parallel algorithms for intra prediction of luminance component and chrominance component, and design different parallel algorithms for integer DCT transform and inverse transform. Design parallel algorithm based on butterfly fast algorithm; design parallel algorithm for quantization and inverse quantization; optimize parallel algorithm for intra prediction. In addition, this paper designs a fast algorithm of intra prediction which gradually reduces the scope of pattern search. The computational complexity is reduced by more than half under the condition that the loss of prediction precision is small. A parallel algorithm is designed for the fast algorithm. In this paper, all parallel algorithms are programmed on the CPU GPU heterogeneous platform with CUDA language, and a large number of experiments are carried out using high-definition video sequences. Compared with the original algorithm, the in-frame prediction parallel algorithm in this paper has a speedup of 5.6 times on the premise of guaranteeing image quality, and compared with the original algorithm, the fast intra prediction parallel algorithm does not change the image quality. The acceleration ratio can reach 8.5 times.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN919.81

【参考文献】