AVS编码器全零块判决算法研究及AVX指令优化

发布时间：2018-05-14 04:42

本文选题：AVS + 帧间模式选择　；参考：《西安电子科技大学》2014年硕士论文

【摘要】：迈进新世纪,为了有效地解决海量视频数据在传输与存储中所带来的困难,视频压缩技术得到了迅猛地发展。AVS(Audio Video coding Standard)作为我国独立设计与制定的视频编解码标准,它的出现打破了国外在视频编解码领域的垄断地位。在AVS中应用了很多创新技术,使得它拥有很好的压缩性能。然而它优越的压缩性能是用牺牲复杂度来换取的。因此,在不影响视频主客观质量的条件下,怎样去降低AVS的复杂度是当前一个研究热点。本文对开源x AVS编码器进行研究,其研究目的是从算法和指令优化两个方面在原始的编码器基础上进一步提升编码速率。针对这一目的,本文完成的工作主要包括了以下的内容:首先,根据整数变换与量化的特点,结合全零块预判决的本质思想,推导出了基于变换量化特点的全零块预判决条件。同时依据图像残差系数服从拉普拉斯统计分布的理论,推导出了基于拉普拉斯统计模型的全零块预判决条件。结合上述所得到的两个预判决条件,给出了一种基于混合模型的全零块预判决算法。实验结果表明此算法相比于已有的算法全零块检出率平均提高20%左右。然后,在基于混合模型的全零块预判决算法基础上,给出了一种基于全零块帧间模式选择提早结束算法。此算法能够提前终止帧间模式选择,跳过变换与量化操作,节省编码时间。实验结果表明此算法编码速率在原始编码器的基础上平均提高了27%左右,而峰值信噪比平均下降仅为0.05dB,传输码率平均上升仅为2.04%。最后,分析并统计了编码器各个关键模块耗时的情况。在支持AVX(Advanced Vector Extensions)多媒体指令集的平台上,分别对耗时较大和密集计算的模块进行优化,如计算SAD(Sum of Absolute Difference)、分像素插值、帧内预测以及变换量化等。与C代码模块相比,指令优化后的模块实现了单指令多数据流的处理方式。通过采用1080P视频序列测试,与原始的编码器进行性能比较,使用峰值信噪比、码率以及编码速率三个指标对比分析优化性能,其结果表明在峰值信噪比和传输码率保持不变的情况下,指令优化后的编码速率要比原始的编码器平均提高3.11倍。结合AVX指令集与基于全零块帧间模式选择提早结束算法,实验结果表明优化后的编码速率在原始编码器的基础上平均提高了3.37倍。
[Abstract]:In the new century, in order to effectively solve the difficulties caused by the mass video data transmission and storage, the video compression technology has been developed rapidly. AVSU Audio Video coding Standard) is a video coding and decoding standard designed and formulated independently in China. Its appearance broke the foreign monopoly in the field of video coding and decoding. Many innovative technologies have been applied in AVS, which makes it have good compression performance. However, its superior compression performance is achieved at the expense of complexity. Therefore, how to reduce the complexity of AVS without affecting the subjective and objective quality of video is a hot topic. In this paper, the open source x AVS encoder is studied. The purpose of the research is to improve the coding rate on the basis of the original encoder from two aspects: algorithm and instruction optimization. For this purpose, the work accomplished in this paper mainly includes the following contents: firstly, according to the characteristics of integer transformation and quantization, combined with the essential idea of all-zero block pre-decision, the condition of all-zero-block pre-decision based on the characteristic of transform quantization is derived. At the same time, according to the theory of image residual coefficient from Laplacian statistical distribution, the full zero block pre-decision condition based on Laplacian statistical model is derived. Combined with the above two pre-decision conditions, an all-zero-block pre-decision algorithm based on mixed model is presented. The experimental results show that the detection rate of this algorithm is about 20% higher than that of the existing algorithms. Then, on the basis of the all-zero block pre-decision algorithm based on hybrid model, an early termination algorithm based on all-zero-block inter-frame mode selection is proposed. This algorithm can terminate the mode selection between frames ahead of time, skip the operation of transformation and quantization, and save coding time. The experimental results show that the coding rate of the proposed algorithm is increased by about 27% on the basis of the original encoder, while the peak signal-to-noise ratio (PSNR) is only 0.05dB, and the average transmission rate is only 2.04b. Finally, the time consuming of each key module of encoder is analyzed and counted. On the platform that supports AVX(Advanced Vector extensions multimedia instruction set, the modules which are time-consuming and intensive are optimized, such as computing SAD(Sum of Absolute difference, sub-pixel interpolation, intra-frame prediction and transform quantization, etc. Compared with C code module, the optimized instruction module realizes the processing of single instruction and multi data flow. By using 1080P video sequence test, compared with the original encoder, using the peak signal to noise ratio (PSNR), bit rate and coding rate to compare and analyze the optimized performance. The results show that the optimized coding rate is 3.11 times higher than that of the original encoder under the condition that the PSNR and the transmission rate remain the same. Combining the AVX instruction set with the early termination algorithm based on the all-zero block inter-frame mode, the experimental results show that the optimized coding rate is 3.37 times higher than that of the original encoder.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN919.81

【参考文献】