基于长短时记忆网络的H.266编码帧内预测优化算法  

Optimization of Intra Prediction Algorithm Based on Long Short-Term Memory Network for H.266 Encoder

作　　者：付青瑫 卯福启[1] FU Qingtao;MAO Fuqi(Col.of Information,North China Univ.of Tech.,100144,Beijing,China)

机构地区：[1]北方工业大学信息学院,北京100144

出　　处：《北方工业大学学报》2020年第2期52-56,69,共6页Journal of North China University of Technology

基　　金：教育部人文社会科学研究一般项目“基于深度学习的海量视频档案知识发现技术研究”(20YJA870014).

摘　　要：在H.266标准帧内预测的角度模式中,编码器根据邻近参考像素使用多抽头帧内插值滤波器生成预测块,预测残差偏大,降低了编码效率.论文提出一种针对水平和垂直模式的优化算法,利用长短时记忆网络表达相邻像素间的空域相关性,对预测残差进行二次预测,补偿标准线性预测过程,提高预测精度.实验结果表明,相比于原始的参考模型VTM2.0,结合长短时记忆网络的帧内预测算法可以使BD-rate降低0.34%,提高了编码效率.In the angle mode of intra prediction in H.266 standard, the encoder uses a multi-tap frame interpolation filter to generate prediction blocks based on neighboring reference pixels, which leads to large prediction residual and reduces the coding efficiency. A new optimization algorithm is proposed in this paper for the horizontal and vertical modes defined in intra prediction. Long Short-Term Memory(LSTM) network is used to describe the spatial correlation between adjacent pixels and to make a compensation prediction of the prediction residual to compensate the standard linear prediction process. The experimental results show that, compared with the original reference model VTM2.0, the intra prediction algorithm combined with the LSTM network can reduce the BD-rate by 0.34% and improve the coding efficiency.

关 键 词：LSTM网络 帧内预测 深度学习 H.266/VVC标准 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]