基于统计建模的VVC快速码率估计算法

Statistical modeling based fast rate estimation algorithm for VVC

为降低新一代通用视频编码(versatile video coding,VVC)标准率失真优化过程的编码复杂度,提出一种基于统计建模的快速码率估计算法。首先,算法充分考虑依赖性量化(dependent quantization,DQ)的量化行为和熵编码中的上下文依赖,提出可以准确刻画编码过程中上下文状态迁移的码率特征,初步预估变换单元(transform unit,TU)中部分语法元素的码率;其次,基于系数分布特性,定义系数混乱度特征和稀疏度特征来区分系数分布差异带来的码率影响,并构建TU级码率模型;最后,算法根据码率构成特性将大尺寸TU和小尺寸TU分开建模实现更精准的码率预估。通过统计方式对大量样本进行回归训练,得到最终的线性码率模型,并应用于VVC的模式决策中。实验结果表明,所提出算法在随机访问(random access,RA)配置下,可以实现16.289%的复杂度降低,而码率变化率(Bjontegaard delta bit rate,BD-BR)仅增加1.567%。

To reduce the coding complexity of the rate-distortion optimization process of the latest video coding standard ver-satile video coding(VVC),a fast rate estimation model based on statistical modeling was proposed.Firstly,the quantization behavior in dependent quantization(DQ)and the context dependency in entropy coding were fully considered.Features that could describe context state transition in the coding process were proposed to estimate rate of some synatax elements in a TU preliminarily.Secondly,coefficient chaos and sparsity features were proposed to distinguish the influence of coefficient distri-bution difference on the rate cost based on the coefficient distribution characteristics which built a TU level rate model.Finally,large-size transform unit(TU)and small-size TU was modeling respectively according to the rate composition character to achieve more accurate rate estimation.A large number of parameters were trained by regression model through statistical me-thods,and the final linear rate model was obtained which was applied to the mode decision.Experimental results show that the proposed algorithm can achieve 16.289%complexity reduction with 1.567%BD-BR increase for RA configuration.