摘要
针对新一代多用途视频编码(versatile video coding,VVC)标准相比上一代高效视频编码(high efficiency video coding,HEVC)采用了更多数目的时空预测模式,为相邻编码帧带来了更强的帧间相关性的问题,基于深度增强学习方法提出了一种适用于VVC编码器的码率控制算法。首先选择合适的模型输入信息,包括帧间相关信息、分层编码结构信息和视频内容信息等;其次利用上述信息,结合长短期记忆(long short-term memory,LSTM)神经网络和增强学习方法,构建基于深度增强学习的帧间量化参数预测模型,以优化VVC编码器的码率控制过程;最后验证所提出算法的性能,将所提出算法在VTM 5.1平台实现,并与VVC源编码器进行性能对比。测试结果表明,在相同码率条件下,所提出算法相比于VVC源编码器,实现了BDBR平均节省1.81%和BDPSNR提升0.14 dB。
Aiming at the previous generation of high efficiency video coding(HEVC),the new generation of versatile video coding(VVC)adopts more spatiotemporal prediction modes,which brings stronger inter frame correlation to adjacent coded frames.Therefore,a rate control algorithm for VVC encoder was proposed based on deep reinforcement learning.Firstly,the appropriate model input information was selected,including inter frame correlation information,hierarchical coding structure information and video content information,etc.Secondly,the above information was combined with long short-term memory(LSTM)neural network.Finally,the performance of the proposed algorithm was verified,and the proposed algorithm was implemented on VTM 5.1 platform and compared with VVC source encoder.The test results show that the proposed algorithm can save BDBR by 1.81%and improve BDPSNR by 0.14 dB compared with VVC source encoder at the same bit rate.
作者
徐艺文
刘航
黄景泉
赵铁松
XU Yiwen;LIU Hang;HUANG Jingquan;ZHAO Tiesong(College of Physics and Information Engineering, Fuzhou University, Fuzhou 350108, China)
出处
《中国科技论文》
CAS
北大核心
2021年第7期748-753,共6页
China Sciencepaper
基金
国家自然科学基金资助项目(61671152)
福建省自然科学基金资助项目(2019J01222)。
关键词
深度增强学习
多用途视频编码
码率控制
帧间相关性
量化参数
deep reinforcement learning
versatile video coding(VVC)
rate control
inter-frame dependency
quantization parameter
