摘要
在R-λ帧内码控中,提出基于卷积神经网络(Convolutional Neural Networks,CNN)的最佳比特分配和最优拉格朗日因子λ选择。首先,探索编码树单元(Coding Tree Unit,CTU)的码率与失真(Rate-Distortion,R-D)及码率与拉格朗日因子λ(Rate-λ,R-λ)的关系特性,设计具有四输出的CNN预测R-D和R-λ曲线的关键参数;然后,建立帧级λ和目标码率的优化方程,反演得到最佳CTU码率分配;最后,根据CTU码率分配和先知的R-λ曲线,得到最优CTU级λ。实验表明,算法在保持4.76%控制精度下,比VTM13.0默认码控算法提高0.31 dB的编码质量。
Optimal bit allocation and optimal Lagrange factorλselection based on Convolutional Neural Networks(CNN)are proposed in R-λintra-frame code control.Firstly,we explore the properties of the coding tree unit(CTU)in terms of the relationship between the code rate and distortion(R-D)and the code rate and the Lagrange factorλ(Rate-λ,R-λ),and design the key parameters of the CNN with four outputs for predicting the R-D and R-λcurves;and then,we establish the optimization of the frame-levelλand target code rate equations and invert them to obtain the optimal CTU code rate allocation;finally,the optimal CTU levelλis obtained based on the CTU code rate allocation and the prophet′s R-λcurve.experiments show that the algorithm improves the coding quality by 0.31 dB over the VTM13.0 default code-control algorithm while maintaining 4.76%control accuracy.
作者
林湧
杨郑龙
罗亦茜
刘欣昱
LIN Yong;YANG Zhenglong;LUO Yixi;LIU Xinyu(School of Urban Railway Transportation,Shanghai University of Engineering Science,Shanghai 201620,China)
出处
《智能计算机与应用》
2024年第4期60-67,共8页
Intelligent Computer and Applications
基金
张江国家自主创新示范区专项发展资金重点项目(ZJ2020-ZD-009)。
关键词
卷积神经网络
帧内码率控制
码率失真和R-λ关系
率失真优化
convolutional neural network
intra frame rate control
rate distortion and R-λrelationship
rate distortion optimization
