有记忆信源率失真模型及高并行度HEVC编码算法优化方法

In the applications such as video surveillance with massive users, specific hardware is the most cost-effective implementation way for HEVC video coder. Hardware oriented algorithm design suffers from multiple challenges: serial dependency, throughput burden, complexity fluctuation, and global algorithm optimization methodology. Building resolvable source model with high precision is one efficient way to address these challenges. This project starts from the intrinsic characteristics of the optimal quantization with dynamic programming and high-order entropy coding of non-memoryless source, and will focus on the fundamental problems on source model and optimization methodology of video coding algorithm. The researches mainly include pipelined optimized quantization, non-memoryless rate distortion model and global algorithm optimization methodology. The innovations of this proposal are as follows. First, we will explore accurate probability density function with multiple segments for DCT transform coefficients, and then propose quantization algorithm with high parallelism based on intelligent prequantization decision under the Bayes theory guidance. Second, we will explore the high-order context entropy coding and the inner relationship between soft-decision optimal quantization and the ratio of the zero-quantized coefficients, and then propose the multiple level rate distortion models for nonmemoryless source Lagrange multiplier and optimization framework for video coding algorithms, using the analytical and heuristic modeling ways. Third, we will explore the mechanisms of inter-frame distortion propagation and intra-frame rate propagation due to high-order context coding, and then propose the spatiotemporally dependent rate distortion models, and then explore the global algorithm optimization methods for optimized video coder design. This project will offer theoretical support for algorithm optimization under the constraints of high parallelism and a certain degree of complexity peak.

在视频监控等海量用户应用中，芯片是HEVC编码高性价比实现方案，其中算法优化面临下述挑战: 高串行依赖、高吞吐压力、复杂度不均衡及算法全局优化；高精度可解析信源模型是应对挑战的有效方法。本项目从动态规划最优量化和高阶熵编码特性入手，针对模型和方法共性问题开展研究，以期提出有记忆信源率失真模型和算法优化方法。研究内容包括：高并行度最优量化算法、有记忆信源模型及算法优化方法、空时域依赖全局优化方法。研究创新如下：首先，基于分段逼近系数模型及系数间相关性，采用贝叶斯判决实现最优量化结果预判，提出高并行度量化算法；其次，根据软判决量化特点准确度量零系数比例，基于此采用启发式分析法提出多层次率失真模型，并改进拉格朗日系数模型及优化方法；再次，定量分析帧间失真传递、高阶熵编码帧内码率传递，提出基于原始图像预分析的空时域依赖率失真模型，改进全局算法优化方法。本研究可为高并行度算法优化提供模型和方法支持。

在视频监控等海量摄像头应用中，专用芯片是HEVC编码高性价比实现方案，设计高性能HEVC编码芯片的关键是设计率失真复杂度优化的算法，算法优化面临几个挑战: 高串行依赖、高吞吐压力、复杂度不均衡及算法全局优化。构建高精度可解析信源率失真模型是应对这些挑战的有效方法。本项目从动态规划最优量化和高阶熵编码特性入手，针对模型和方法共性问题开展了系统研究，提出有记忆信源率失真模型和算法优化方法。研究内容包括：（1）高并行度率失真优化量化算法；（2）有记忆信源率失真模型及率失真算法优化方法；（3）空时域失真码率依赖的全局编码优化方法。..重要研究成果包括： 发表SCI检索论文10篇，其他重要会议论文11篇，授权发明专利4项，正在公开的专利9项，其他一级期刊论文3篇。.创新研究成果如下：首先，基于分段逼近系数模型及系数间相关性，采用贝叶斯判决实现最优量化结果预判，提出高并行度率失真优化量化算法；其次，根据软判决量化特点确定了高准确度的零系数比例模型，基于此采用启发式分析法提出多层次率失真模型，并改进拉格朗日系数模型及优化方法；再次，定量分析帧间失真传递、高阶熵编码帧内码率传递，提出基于原始图像预分析的空时域依赖率失真模型，改进全局算法优化方法。..本研究的科学意义: 本项目研究有望为高阶上下文编码码率模型、高并行度量化、模式选择、码率控制算法优化提供理论依据，可为HEVC和AVS2编码芯片实现算法优化提供信源理论支撑。